SkillNEO：智能体技能的非精英进化优化

# SkillNEO: Non-Elitist Evolutionary Optimization of Agent Skills
Source: [https://arxiv.org/html/2610.20926](https://arxiv.org/html/2610.20926)
Enze Ma,Chunyu Miao,Yankai Chen,Jie Yang,Liancheng Fang,Kening Zheng,Hanrong Zhang,Muyao Xu,Xinle Yu,Chengze Li,Wei\-Chieh Huang,Ziwei Ren,Yunhao Du,Zhixiang Shen,Xue Liu,Philip S\. Yu,Zhen WangUniversity of Illinois at ChicagoMohamed bin Zayed University of Artificial IntelligenceUniversity of California, San DiegoUniversity of California, Los AngelesIndependent ResearcherUniversity of Texas at Arlington

###### Abstract

Agent skills turn procedural knowledge into reusable artifacts, allowing LLM agents to accumulate and refine expertise through experience\.Building on this, skill evolution uses execution feedback to improve skills over successive revisions\. However, improvement may require several interdependent edits, so a low\-reward version can still be worth developing\.Requiring immediate reward gains can therefore discard useful partial repairs before their benefits emerge\. Non\-elitist evolution allows new versions to receive further edits even when they score below earlier ones\. For agent skills, this opportunity also calls for evidence of what to preserve and what to repair in the next edit\. We introduceSkillNEO\(Non\-Elitist Evolutionary Optimization for Agent Skills\), which separates the choice of a skill version to use from the choice of versions to keep improving\. By comparing how related versions carry out the same task, it identifies working steps to preserve and remaining failures to fix\. It can also return to earlier versions and revise them using what later attempts reveal\. Across 86 SkillsBench tasks spanning software engineering, office work, and six other domains,SkillNEOraises GPT\-5\.4 Mini’s average reward from 49\.1% to 66\.4%, bringing it into GPT\-5\.5 and GPT\-5\.6 Luna’s original\-skill performance range\. Mini executes tasks, proposes revisions, and edits skills, with final evaluation using the same task instances\. Gains extend across all eight domains\. The same skills also improve stronger models without further editing\. On the 41 tasks with modified skills, reuse improves Luna by 10\.5 percentage points and places GPT\-5\.5 10\.6 points above its published original\-skill reference on those tasks\. Execution records show that these skills can benefit stronger models even when Mini still fails some task requirements\. Recorded searches also show lower\-scoring versions leading to the skills ultimately selected for use\. Together, these findings show that skill evolution can turn a weaker model’s partial progress into reusable procedural knowledge for improving more capable agents\.

## 1Introduction

LLM agents increasingly perform tasks that require procedural expertise, including how to use domain tools, coordinate dependent steps, and verify intermediate results\([Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1);[Zhang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib4)\)\. Agent skills make this expertise reusable through natural\-language guidance, executable programs, and workflows\([Wang et al\., 2023](https://arxiv.org/html/2610.20926#bib.bib17);[Wang et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib28);[Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1)\)\. Such reusable procedures already support web interaction, memory management, and even skill learning itself\([Zheng et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib23);[Zhang et al\., 2026c](https://arxiv.org/html/2610.20926#bib.bib25);[Zhang et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib12)\)\. Beyond guiding execution, these artifacts can also retain and refine lessons from earlier attempts\([Wang et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib28);[Zhou et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib26);[Ni et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib22)\)\. Building on this, skill evolution revises instructions, scripts, and resources using execution feedback, then selects versions for further development by task performance\([Alzubi et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib7);[Wang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib27)\)\.

However, a skill’s current task reward can conceal partial progress within its procedure\. We call the skills provided for a task and their associated files a*skill bundle*\. Because the instructions, scripts, and supporting resources within a bundle can be coupled\([Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1);[Zhang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib4)\), completing a repair may require several interdependent edits\. For example, a revision can correct input processing while leaving output construction incomplete, so the whole\-task reward remains unchanged or even declines\. Yet requiring improvement over the parent\([Wang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib27);[Shastri et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib2)\)or retaining the highest\-scoring candidates across generations\([Alzubi et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib7);[Sun et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib39)\)can exclude this partial repair from further editing\. The challenge is therefore to distinguish a version’s readiness for deployment from its value as a starting point for further repair\.

Non\-elitist evolutionary selection provides a concrete way to preserve these opportunities for further repair \([Beyer and Schwefel, 2002](https://arxiv.org/html/2610.20926#bib.bib33);[Oliveto et al\., 2018](https://arxiv.org/html/2610.20926#bib.bib50)\)\. In offspring\-only replacement, unchanged parents do not compete for the next generation’s places, while candidates still compete on measured performance within their generation\. A partially repaired child can therefore receive further edits even when its parent or an earlier version scores higher\. For skill evolution, this permits a sequence of revisions to continue before a reward gain emerges, while the best available version remains deployed\. Further development requires guidance on how to complete each repair\. Task reward alone cannot identify working steps or unresolved failures\([Agrawal et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib9)\)\.

Figure 1:Non\-elitist evolution allows weaker revisions to lead to better skills\.\(a\) Single\-direction refinement discards alternative revisions, potentially losing useful partial repairs\. \(b\) SkillNEO develops and revisits such revisions, using execution\-trace comparisons to guide further repairs that can yield a new best skill\. Statessis\_\{i\}pair skill bundles with their execution records\. The dark\-blue node marks the new deployment incumbent; crossed\-out nodes in \(a\) denote discarded revisions\. Solid tree edges denote edits; in \(b\), the dashed orange arrow denotes reselection\.We realize this opportunity inSkillNEO, a framework for*Non\-Elitist Evolutionary Optimization for Agent Skills*that guides non\-elitist skill development through comparative execution evidence \(Figure[1](https://arxiv.org/html/2610.20926#S1.F1)\)\. Each candidate pairs a skill bundle with its execution history, preserving both a starting point for further edits and the evidence needed to direct them\. We organize comparisons of these executions along the revision tree, building on execution\-guided skill diagnosis and refinement\([Gao et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib3);[Ni et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib22)\)\. Parent\-child comparisons relate behavioral changes to edits, while sibling comparisons contrast alternative repairs from a shared starting point\. These comparisons reveal working procedures and unresolved failures, guiding coordinated changes to instructions, scripts, and resources, including reuse across branches\. Archive revisiting extends this process to earlier versions, allowing them to be revised using evidence gathered after their creation\.

We evaluate task\-conditioned skill development on 86 SkillsBench tasks, with final evaluation using the same task instances\. With GPT\-5\.4 Mini\([OpenAI, 2026b](https://arxiv.org/html/2610.20926#bib.bib31)\)performing task execution, proposal generation, and editing,SkillNEOraises its average reward from 49\.1% to 66\.4%, within GPT\-5\.5 and GPT\-5\.6 Luna’s original\-skill performance range\. Recorded searches show lower\-scoring versions becoming ancestors of deployed bundles and supplying useful procedures to other branches\. Beyond Mini’s own gains, the frozen skills deliver substantial improvements when reused by stronger executors\. On the 41 modified tasks, reuse improves Luna by 10\.5 percentage points; GPT\-5\.5 exceeds its published original\-skill reference on these tasks by 10\.6 points\. Together, these results suggest that skill evolution can accumulate procedural knowledge across revisions and extend the benefits of one model’s experience to more capable agents\.

## 2Related Work

Agent skills and reusable knowledge\.Agent skills take forms such as natural\-language guidance, executable programs, workflows, and learned policies\([Wang et al\., 2023](https://arxiv.org/html/2610.20926#bib.bib17);[Wang et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib28);[Jiang et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib21)\)\. These skills support web operations\([Zheng et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib23)\), shared interfaces across websites\([Yu et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib24)\), and further skill development\([Zhang et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib12)\)\. Previous studies also show that skills can transfer across models\([Zheng et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib23);[Zhang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib4);[Yang et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib19);[Tan et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib8);[Xiong et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib37)\)\. Related agent research covers memory management and auditing\([Zhang et al\., 2026c](https://arxiv.org/html/2610.20926#bib.bib25);[Ma et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib42)\), as well as scientific workflows and their evaluation\([Gao et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib47);[Wang et al\., 2026c](https://arxiv.org/html/2610.20926#bib.bib46)\)\. Skill and agent evolution refine reusable artifacts over successive rounds using execution feedback\([Zhou et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib26);[Tan et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib8);[Hu et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib45)\)\. ACE refines context as an evolving playbook\([Zhang et al\., 2026e](https://arxiv.org/html/2610.20926#bib.bib15)\), while Meta Context Engineering evolves the skills used to construct and update context\([Ye et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib14)\)\. We study how to choose versions for continued development and use their execution records to guide revisions\.

Search and candidate selection\.Search methods explore alternative candidates for reasoning\([Hao et al\., 2023](https://arxiv.org/html/2610.20926#bib.bib10);[Hao et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib5);[Wang et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib49)\)and scientific discovery\([Liu et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib44);[Yu et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib43)\)\. Prompt optimizers use feedback and beam search\([Pryzant et al\., 2023](https://arxiv.org/html/2610.20926#bib.bib11);[Singla et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib48)\), error\-guided tree search\([Wang et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib6)\), prompt\-score histories\([Yang et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib35)\), or evolving mutation prompts\([Fernando et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib34)\)\. NPO reports that single\-lineage search is competitive in its evaluated settings\([Chang and Chen, 2026](https://arxiv.org/html/2610.20926#bib.bib36)\)\.

Candidate retention determines which skill versions receive further edits\. Evolutionary comma selection replaces parents with offspring, allowing the active population’s best fitness to decline across generations\([Beyer and Schwefel, 2002](https://arxiv.org/html/2610.20926#bib.bib33)\)\. Skill\-evolution methods differ in the conditions they impose on continued development\. Once its validation\-ranked frontier is full, EvoSkill admits only candidates that outperform its lowest\-scoring member\([Alzubi et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib7)\)\. GraphSkillEvo uses mutation and crossover on graph\-structured skills, retaining the highest\-scoring candidates from the combined parent\-child pool\([Sun et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib39)\)\. MetaSkill\-Evolve retains failed revisions as inspiration but requires positive parent\-relative gain for future parent eligibility\([Wang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib27)\)\.

Several methods preserve alternatives through tree search, multiple populations, or an archive\. BONSAI revisits weaker branches through an evolvability\-guided tree policy, supports grafting, and separates search from deployment; ordinary rewrites must improve on their parent on the reflection minibatch\([Shastri et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib2)\)\. DIVE retains parents alongside children, targets lower\-performing skills for exploratory revision, evolves independent populations, and selects complementary skills for inference\([Xiong et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib37)\)\. DGM evolves coding agents by sampling archived implementations for further self\-modification\([Zhang et al\., 2026d](https://arxiv.org/html/2610.20926#bib.bib41)\)\. Our ordinary update selects parents only from eligible, fully measured children in the current generation\. Previous parents do not compete, and children need no parent\-relative gain\. Eligible earlier bundles can return to the parent pool through a separate archive\-revisiting rule\.

Execution\-guided skill refinement\.Execution feedback can supply both reusable guidance and evidence for its revision\. AWM abstracts workflows from execution histories\([Wang et al\., 2025](https://arxiv.org/html/2610.20926#bib.bib28)\), while SkillRL co\-evolves an experience\-derived skill library with a reinforcement\-learned policy\([Xia et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib13)\)\. OpenClaw\-Skill searches skill compositions across subtasks\([Lin et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib20)\); our tree records alternative versions of a complete bundle\. Reflexion stores reflections for later attempts\([Shinn et al\., 2023](https://arxiv.org/html/2610.20926#bib.bib16)\), while ExpeL compares successful and failed executions to update reusable insights\([Zhao et al\., 2024](https://arxiv.org/html/2610.20926#bib.bib18)\)\. GEPA reflects on trajectories and merges prompts across lineages\([Agrawal et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib9)\); EvoSkill and MetaSkill\-Evolve reuse proposal histories or retrieved inspirations\([Alzubi et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib7);[Wang et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib27)\)\. Trace2Skill consolidates trajectory\-level patches into a portable skill directory\([Ni et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib22)\); SkillAudit compares executions with and without a skill to guide revision\([Gao et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib3)\)\. Branch2Skill compares shared\-prefix reasoning paths for local supervision\([Ren et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib38)\); SkillLift learns an oracle\-aligned rubric to guide revisions between costly evaluations\([Kang and Wen, 2026](https://arxiv.org/html/2610.20926#bib.bib40)\)\.SkillNEOlinks executions to saved bundle versions, comparing repeated runs of the same version and parent\-child or sibling executions\. This evidence guides joint edits to instructions and supporting files, including reuse of working procedures from versions with unresolved failures\. Later executions can guide edits to saved bundles\.

## 3SkillNEO for Non\-Elitist Evolution of Agent Skills

![Refer to caption](https://arxiv.org/html/2610.20926v1/skillneo_method_artwork.png)Figure 2:SkillNEO separates choosing a skill to use from choosing versions to improve\.A version can receive further edits \(green\) even when it scores below the version selected for use \(blue\)\. To guide these edits, the Proposer compares saved task executions to identify steps that work and failures that need repair\. The Skill Editor applies the proposed changes to instructions, scripts, and resources\. Edit Validation checks the revised files before the Executor runs the task with the new version\. The resulting scores inform selection, and execution records guide the next round\. Mini fills every model\-based role in our experiments\.SkillNEOmaintains a deployed bundle while developing alternatives that may need further repair \(Figure[2](https://arxiv.org/html/2610.20926#S3.F2)\)\. Two decisions govern this process\. Deployment determines whether a measured candidate should replace the incumbent; development determines which versions can receive another edit\. Non\-elitist selection keeps these decisions separate\. Comparisons of related executions guide these revisions by identifying working procedures to preserve and unresolved failures to address\.

### 3\.1Problem Formulation

Given a task and an initial skill bundle, we seek revisions that improve a frozen agent’s expected task reward\. Letttdenote a task with a fixed instruction, execution environment, and verifierVtV\_\{t\}\. SkillsBench supplies each task with a set of relevant skills\([Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1)\)\. Each skill contains its ownSKILL\.mddocument and optional executable tools and resources\. A bundleLLstores a version of this entire set, andSkillNEOevolves each task’s bundle independently\. A frozen agentAmA\_\{m\}with modelmmexecutesttusingLL, producing a trajectoryτ\\tauand an outcomeyycontaining the submitted artifacts and final environment state\. The verifier assigns rewardVt​\(y\)∈\[0,1\]V\_\{t\}\(y\)\\in\[0,1\]\. We defineJm​\(L,t\)J\_\{m\}\(L;t\)as this agent’s expected reward on taskttwith bundleLL\.

\(τ,y\)∼Am​\(t,L\),Jm​\(L,t\)=𝔼\(τ,y\)∼Am​\(t,L\)​\[Vt​\(y\)\]\.\(\\tau,y\)\\sim A\_\{m\}\(t,L\),\\qquad J\_\{m\}\(L;t\)=\\mathbb\{E\}\_\{\(\\tau,y\)\\sim A\_\{m\}\(t,L\)\}\[V\_\{t\}\(y\)\]\.\(1\)The expectation is over execution variability for the same task, model, and bundle\. An execution recorde=\(τ,y,r,d\)e=\(\\tau,y,r,d\)contains the trajectory, outcome, rewardr=Vt​\(y\)r=V\_\{t\}\(y\), and filtered verifier findingsdd\. Traces and findings guide revisions; rewards measure task performance\. Controlled bundle comparisons use the same task instruction, execution environment, verifier, and model configuration\.

Starting from the original bundleLt0L\_\{t\}^\{0\}, edits define a space of reachable bundlesℒt\\mathcal\{L\}\_\{t\}\. An edita=\(target,operator,specification\)a=\(\\text\{target\},\\text\{operator\},\\text\{specification\}\)producesL′=Apply⁡\(L,a\)L^\{\\prime\}=\\operatorname\{Apply\}\(L,a\)and may revise instructions, executable code, or resources within the bundle\. The ideal objective and the output of a finite search are

maxL∈ℒt⁡Jm​\(L,t\),L^t=𝒮κ​\(t,Lt0,m\),\\max\_\{L\\in\\mathcal\{L\}\_\{t\}\}J\_\{m\}\(L;t\),\\qquad\\widehat\{L\}\_\{t\}=\\mathcal\{S\}\_\{\\kappa\}\(t,L\_\{t\}^\{0\},m\),\(2\)where𝒮κ\\mathcal\{S\}\_\{\\kappa\}denotes the search procedure andκ\\kappaspecifies its generation limit, candidate width, and evaluation repeats\. We estimateJm​\(L,t\)J\_\{m\}\(L;t\)from a finite number of executions, retaining evaluated bundles and their execution records to guide subsequent edits\. Once development ends for tasktt, we freeze its selected deployment bundleL^t\\widehat\{L\}\_\{t\}\. Running this procedure independently across tasks𝒟\\mathcal\{D\}yields the task\-indexed assignment𝐋^=\(L^t\)t∈𝒟\\widehat\{\\mathbf\{L\}\}=\(\\widehat\{L\}\_\{t\}\)\_\{t\\in\\mathcal\{D\}\}\. Cross\-model reuse then evaluates these same task\-bundle pairs with another frozen model, without repeating skill development for the receiver\.

### 3\.2Non\-Elitist Skill Development

We retain both the bundle and its execution history so that a saved version can serve as a starting point for further repair\. For tasktt, a rooted tree𝒯t=\(𝒱t,𝒜t\)\\mathcal\{T\}\_\{t\}=\(\\mathcal\{V\}\_\{t\},\\mathcal\{A\}\_\{t\}\)records revisions\. Nodevvstoressv=\(Lv,Ev\)s\_\{v\}=\(L\_\{v\},E\_\{v\}\), whereLvL\_\{v\}is a complete bundle andEvE\_\{v\}contains its execution records\. An editaua\_\{u\}creates childuufrom a saved primary parentp⁡\(u\)p\(u\),

Lu=Apply⁡\(Lp⁡\(u\),au\),\(p⁡\(u\),u\)∈𝒜t\.L\_\{u\}=\\operatorname\{Apply\}\(L\_\{p\(u\)\},a\_\{u\}\),\\qquad\(p\(u\),u\)\\in\\mathcal\{A\}\_\{t\}\.\(3\)The root storesL0≡Lt0L\_\{0\}\\equiv L\_\{t\}^\{0\}\. A graft can reuse another version’s procedure while retaining one primary parent edge\. At the end of generationgg,hgh\_\{g\}indexes the deployment incumbent, andBgB\_\{g\}contains the versions eligible to parent edits in generationg\+1g\+1\.

The ordinary development update selects parents from the current generation’s children\. With parent\-pool capacitykk, we screen admitted children and promote up tokkfor repeated evaluation\. LetCgC\_\{g\}contain the promoted children with complete evaluations and available bundles\. We form the ordinary parent set from these children,

Bgord=Topk⁡\(Cg\)\.B\_\{g\}^\{\\mathrm\{ord\}\}=\\operatorname\{Top\}\_\{k\}\(C\_\{g\}\)\.\(4\)Previous parents do not compete for places in this pool\. This offspring\-only replacement follows non\-elitist comma selection\([Beyer and Schwefel, 2002](https://arxiv.org/html/2610.20926#bib.bib33)\)\. Children compete on measured task reward within their generation, without needing to outperform their own parents or enter the historical top\-kk\. A child can therefore receive another edit even when its current reward is too low for deployment\.

This ordinary update supplies the default parent set\. Archive revisiting can replace it with eligible earlier versions under the conditions in Section[3\.4](https://arxiv.org/html/2610.20926#S3.SS4)\. We then add the measured root if fewer than two parents remain and the root is absent\. The resultingBgB\_\{g\}supplies the next round of edits, while the deployment decision separately determineshgh\_\{g\}\.

### 3\.3Comparative Evidence for Skill Revision

Task rewards rank outcomes; the executions behind those outcomes help identify a useful next edit\. We compare repeated runs of one version to examine execution variability, parent\-child runs to relate behavioral changes to an edit, and sibling runs to examine alternative repairs\. Generation one uses repeated root executions\. Later generations add population scores, traces, tool outputs, submitted artifacts, routing records, and edit differences linked to their bundle versions\. This evidence includes screening\-only children outside the parent pool, so a version can inform a revision even when it is ineligible to supply its starting bundle\.

The*Proposer*uses these comparisons to identify unresolved failures and successful procedures worth preserving\. Routing records distinguish failure to consult a skill from failure after consulting it\. The first can call for revising the skill’s visible description; the second calls for examining its procedure and supporting files\. We instruct the Proposer to rank suspected causes by the failed runs they explain and check them against passing executions\. Errors that an agent recovers from do not by themselves justify a revision\. Each proposal identifies a primary parent fromBg−1B\_\{g\-1\}, an intended change, and supporting evidence\. Verifier findings are filtered before entering the proposal context\. The prompt also discourages repeating edit directions previously measured as net losses; this constraint on proposed edits does not change parent eligibility\.

The*Skill Editor*turns a proposal into a complete child bundle while preserving the saved parent\. A repair can require coordinated changes across instructions, scripts, and resources, such as adding a helper together with the instruction that invokes it\. The editor can rewrite, add, remove, or reorganize these files, or graft a procedure from another eligible parent\. We instruct it to preserve unrelated content\. Appendix[B](https://arxiv.org/html/2610.20926#A2)specifies the operators and proposal constraints\.

*Edit Validation*checks operation constraints, referenced files, new scripts, and candidate content before task evaluation\. Fixable failures return to the Skill Editor for repair and rechecking; rejected candidates do not enter evaluation\. Admitted candidates are scored by the benchmark verifierVtV\_\{t\}\. Appendix[B\.2](https://arxiv.org/html/2610.20926#A2.SS2)specifies the checks and repair paths\.

### 3\.4Evaluation, Deployment, and Archive Revisiting

Repeated evaluation determines which revisions are ready for deployment\. The*Executor*runs a task with a bundle and records trajectories, tool outputs, and submitted artifacts\. We estimate bundle performance by averaging verifier rewards over an execution batch \(Equation[6](https://arxiv.org/html/2610.20926#A2.E6)\)\. The root receivesRRinitial executions\. Each generation requestsbbchildren, screens each admitted child once, and promotes up tokkforR−1R\-1additional executions\. The incumbent receivesRRfresh executions\. Deployment ranks promoted children by their mean reward over allRRexecutions, including screening\. Compatible auxiliary scores break official\-score ties at both screening and deployment\. Our experiments use\(b,k,R\)=\(6,3,3\)\(b,k,R\)=\(6,3,3\)\.

We compare the best fully measured child with the incumbent under the configured acceptance rule\. The current\-batch criterion requires an official\-score gain beyond the margin, or an official\-score tie with sufficient auxiliary\-score gain\. Under the strict\-floor configuration, the child’s mean must also exceed the larger pooled official score of the root and incumbent\. A failed comparison leaves the incumbent deployed and produces aHold\. Appendix[B](https://arxiv.org/html/2610.20926#A2)gives the margins, historical checks, and handling of incomplete evaluation records\.

Archive revisiting allows development to return to an earlier branch with evidence collected since that version was created\. When enabled, it can replace the ordinary parent set after aHoldin which the best child’s official mean reward is below the incumbent’s\. Retirement rules determine which saved versions can re\-enter development\. The incumbent remains deployed as the parent set changes\. Search allows up to five generations and stops after two consecutive holds, with additional termination conditions specified in Appendix[B](https://arxiv.org/html/2610.20926#A2)\. Algorithm[1](https://arxiv.org/html/2610.20926#alg1)gives the complete procedure\. The outputL^t=Lhg\\widehat\{L\}\_\{t\}=L\_\{h\_\{g\}\}is frozen before cross\-model reuse\.

## 4Experiments

Table 1:Overall rewards and cross\-model reuse\.Average task reward \(%\) for the 86\-task assignment \(a\) and 41 modified tasks \(b\);Δ\\Deltais in percentage points\.†\\daggerGPT\-5\.4 compares Vals \(87 tasks, OpenHands\) with final evaluation on 86 tasks using Codex\. GPT\-5\.5 uses published, task\-aligned SkillsBench references; Luna uses post\-freeze evaluations of both conditions\. Subsetgains/ties/lossesare 29/12/0 \(Mini\), 15/18/8 \(GPT\-5\.5\), and 11/23/7 \(Luna\), with ties within±0\.05\\pm 0\.05pp\.\(a\) Overall performance

\(b\) 41 modified tasks

### 4\.1Experimental setup

Evaluation scope\.We evaluate how muchSkillNEOimproves Mini’s task performance, then examine recorded search paths to understand how intermediate versions contribute to later revisions\. We also test whether the frozen bundles remain useful to stronger executors\. Development and final evaluation use the same task and input instances\. Appendix[A](https://arxiv.org/html/2610.20926#A1)discusses the limitations\.

Benchmark and metric\.We use 86 professional tasks from SkillsBench 1\.1, spanning eight broad categories\([Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1)\)\. Each task’s supplied skills form its initial bundle, which we develop independently\. For task set𝒟\\mathcal\{D\}, modelmm, and bundle assignment𝐋=\(Lt\)t∈𝒟\\mathbf\{L\}=\(L\_\{t\}\)\_\{t\\in\\mathcal\{D\}\}, repeatqqproduces\(τt,q,yt,q\)∼Am​\(t,Lt\)\(\\tau\_\{t,q\},y\_\{t,q\}\)\\sim A\_\{m\}\(t,L\_\{t\}\)\. We average three rewards per task, then take the mean across tasks,

Sm​\(𝐋\)=100\|𝒟\|​∑t∈𝒟13​∑q=13Vt​\(yt,q\)\.S\_\{m\}\(\\mathbf\{L\}\)=\\frac\{100\}\{\|\\mathcal\{D\}\|\}\\sum\_\{t\\in\\mathcal\{D\}\}\\frac\{1\}\{3\}\\sum\_\{q=1\}^\{3\}V\_\{t\}\(y\_\{t,q\}\)\.\(5\)Average reward is reported as a percentage and retains fractional rewards; it equals success rate only for binary tasks\. Published references retain their documented aggregation\.

Models and skill reuse\.GPT\-5\.4 Mini/Codex\([OpenAI, 2026a](https://arxiv.org/html/2610.20926#bib.bib32)\)at high reasoning performs every LLM call during development, including task execution, proposal generation, and skill editing\. Let𝐋0\\mathbf\{L\}^\{0\}denote the original bundle assignment and𝐋^\\widehat\{\\mathbf\{L\}\}the frozen assignment selected by development\. The latter contains 41 modified bundles and 45 bundles byte\-identical to their originals\. We evaluate these same task\-bundle pairs with GPT\-5\.4/Codex xhigh, GPT\-5\.5/Codex xhigh, and GPT\-5\.6 Luna/Codex max\. Reuse involves no further skill search or feedback to development\.

Final evaluation protocol\.Mini’s 41 modified bundles receive three post\-evolution executions per task, separate from search and selection\. The 45 unchanged bundles share the three original\-skill executions because their original and final configurations are identical\. Mini’s 86\-task score combines these measurements\. Luna evaluates both assignments after skill freezing, yielding 516 valid executions\. Appendix[B](https://arxiv.org/html/2610.20926#A2)gives the reference sources and execution details\.

Table 2:Mini’s rewards across eight domains\.Average task reward \(%\) on the same 86 tasks\. No skill†is a published GPT\-5\.4 Mini/OpenHands[reference](https://github.com/benchflow-ai/skillsbench-trajectories/blob/cb640f73cdea8c67422cae3cc37b913a7c080871/website-data/results-registry.json), reaggregated by domain; Original skills andSkillNEOuse our Mini/Codex high evaluation\.Δ=SkillNEO−Original skills\\Delta=\\text\{\\mbox\{SkillNEO\}\}\-\\text\{Original skills\}, in percentage points\. The Overall row averages all 86 tasks\.Figure 3:Mini improves across all eight domains\.No skill†uses the OpenHands reference in Table[2](https://arxiv.org/html/2610.20926#S4.T2); Original skills andSkillNEOuse our Mini/Codex high evaluation\. Parenthesized gains areSkillNEOminus Original skills \(pp\)\. Domain counts sum to 86\.
### 4\.2Performance across tasks and domains

Mini reaches the performance range of stronger models using original skills\.With Mini in every model\-based development role,SkillNEOraises its average reward across 86 tasks from 49\.1% to 66\.4% \(Table[1](https://arxiv.org/html/2610.20926#S4.T1)a\)\. The final score lies between GPT\-5\.5’s published original\-skill reference of 66\.1% and Luna’s original\-skill score of 67\.3%\. On the 41 modified tasks, fresh post\-evolution evaluation gives Mini 67\.8% against the original\-skill 31\.5%, with 29 gains, 12 ties, and no observed losses\. The other 45 tasks retain their original bundles and Mini scores\.

Mini improves across all eight domains\.Improvements range from 5\.6 pp in industrial / physical tasks to 27\.4 pp in software engineering \(Figure[3](https://arxiv.org/html/2610.20926#S4.F3); Table[2](https://arxiv.org/html/2610.20926#S4.T2)\)\. Both domains start near 48\.3%\.Execution records show how revisions translate task requirements into concrete actions and where those requirements remain unenforced\.

Revised skills help Mini check and satisfy task requirements\.In SimPO, the original runs already produce loss vectors numerically matching passing runs, but use Python 3\.12 despite the Python 3\.10 requirement\. The evolved skill puts environment activation before installation and testing; final executions satisfy the requirement and pass\. In Data\-to\-D3, the revised skill requires interaction checks\. One final execution fixes and retests a bug that clears a row when selected twice\.

However, some revised workflows still leave a requirement unchecked or unenforced\. Adaptive Cruise Control improves only partially; its remaining failed execution checks speed and distance summaries without verifying that emergency mode occurs\. Manufacturing Codebook Normalization’s completed runs use test\-item context for electrical\-fault labels where the revised skill requires component evidence from the raw segment, and still fail codebook checks\. Appendix[D\.1](https://arxiv.org/html/2610.20926#A4.SS1)links the skill revisions to all 24 executions of the original and final bundles\.

Figure 4:Case studies of skill evolution in SkillNEO\.\(a\) A zero\-reward partial repair leads to the next selected skill in Gravitational Wave Detection\. \(b\) Sales Pivot Analysis reuses a sibling’s PDF extraction procedure in another branch\. \(c\) Revisiting an earlier parent produces the final skill in Paratransit\. \(d\) Restarting from the original bundle uses execution evidence collected later in Wyckoff\. Panel \(a\) reports three\-run mean rewards at creation\. In \(c\) and \(d\), solid edges show primary parents, broken arrows show reuse or reselection, and rewards come from independent final Mini evaluations \(Appendices[H](https://arxiv.org/html/2610.20926#A8)and[I](https://arxiv.org/html/2610.20926#A9)\)\.
### 4\.3Search paths and execution evidence

We examine the 41 modified bundles’ development histories to trace how intermediate versions contribute to later skills\. The records show continued repairs, procedure reuse across branches, and edits informed by later executions\. Candidate lettersbthroughfdenote generations 1 through 5;e04, for example, is candidate 04 in generation 4\.

Lower\-scoring versions can lead to better final skills\.Among the 29 tasks with final Mini gains, 22 \(75\.9%\) use a non\-incumbent parent on the final path\. Sixteen paths pass through a candidate that scored below the deployed bundle when it was created\. Across the 41 searches, 163 such lower\-scoring versions receive further edits, and 45 have a later deployed descendant\. For 18 of these 45 versions, the first deployed descendant lies at least two edits away \(Appendix[H](https://arxiv.org/html/2610.20926#A8)\)\.

Zero reward can hide steps that already work\.In Gravitational Wave Detection \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)a\),e04scores zero despite passing eight of nine checks in every execution\. Generation four deployse01but retainse04in a three\-parent pool despite three higher\-scoring saved versions\. The remaining failure concerns the numerical result, with all three runs using the 15 Hz conditioning cutoff for search despite a separate skill specifying 20 Hz\. Childf04consolidates the two skills, placing the 20 Hz search setting alongside the 15 Hz conditioning steps without changing helper code\. All three child executions follow these settings and pass every check\. The search deploysf04at 100\.0% mean reward against the re\-measured incumbent’s 66\.7% \(Appendix[I\.2](https://arxiv.org/html/2610.20926#A9.SS2)\)\.

The figure’s orange curve follows the final skill’s ancestry through the decline toe04and its subsequent repair; the blue curve records the deployment incumbent\. Final evaluation raises Mini from 0\.0% to 100\.0%, while Luna retains 100\.0%\. Appendix[H](https://arxiv.org/html/2610.20926#A8)documents a similar case in Data\-to\-D3, whered03remains in a three\-parent pool despite three higher\-scoring saved versions\.

Comparing runs reveals reusable steps in lower\-scoring skills\.Eight graft edits on six final ancestries draw from donors outside those primary paths\. In Sales Pivot Analysis, the lower\-scoringb05contributes state\-name handling to a sibling branch \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)b\)\. Its two failed runs pass state\-validity checks but fail pivot configuration; a third run succeeds\. Citing that successful execution, thec04proposal selectsb04as parent andb05as donor\. The edit preservesb04’s spreadsheet skill for pivot construction and importsb05’s PDF extraction procedure\. Both passingc04executions use the adopted extraction pattern with full state names\. Although the historical\-score check keepsb04deployed,c04remains eligible for further edits and enters the final ancestry\. The comparison recovers a useful procedure from a bundle with lower reward \(Appendix[I\.4](https://arxiv.org/html/2610.20926#A9.SS4)\)\.

Comparing runs shows what to keep and what to fix next\.Software Dependency Audit requires vulnerabilities in production dependencies, and its original command example already uses that scope\. Revisionb06adds deduplication and checks that the saved CSV matches the scan\-derived records\. Its failed execution verifies 22 findings but still expands the scan to development dependencies; passing executions report three production\-dependency findings\. Checking the report against the scan cannot correct the scan’s scope\. The next revision,c01, makes that scope explicit while preserving the reporting checks\. All three search\-time executions follow the intended scope and pass\. As in Gravitational Wave Detection, the revision places an existing requirement where the agent needs to apply it \(Appendix[I\.1](https://arxiv.org/html/2610.20926#A9.SS1)\)\.

Later runs guide new revisions of earlier skill versions\.Revisiting a saved bundle gives its next edit access to executions unavailable when the bundle was created\.Paratransit illustrates the return to an earlier parent: the search reselectsd06from generation three after evaluating its generation\-four children\.Its generation\-five childf08becomes the final bundle, raising Mini from 0\.0% to 33\.3% \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)c\)\.Wyckoff also shows how later execution evidence informs the new edit \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)d; Appendix[H](https://arxiv.org/html/2610.20926#A8)\): the search returns to the original bundle in generation three and usesb06’s first\-generation execution to make file\-based analysis and structured output requirements explicit ind08\.A new execution follows that guidance, andd08’s generation\-four childe03becomes the final bundle, raising Mini’s independent final reward from 75\.7% to 100\.0%\.

Skill instructions shrink and scripts grow on average\.Across the 41 modified tasks,SKILL\.mdfiles shrink by 35\.0 lines per task on average, support scripts grow by 74\.0 lines, and 18 scripts are added\. Total source length increases by 1\.5% \(Table[13](https://arxiv.org/html/2610.20926#A7.T13)\)\. Erlang SSH \(Appendix[I](https://arxiv.org/html/2610.20926#A9)\) compresses a 485\-line guide into a 31\-line task procedure with perfect final scores for both Mini and Luna\.

### 4\.4Cross\-model reuse

Table 3:Execution behavior with the same final skill bundle\. Models differ in which requirements they satisfy despite receiving identical guidance\. Counts give passes out of three executions; skipped checks are not passes\. For Sales Pivot,Valid pivot objectsrequires the row\-field, aggregation, and cache\-field checks to pass\.Rewardis mean task reward \(%\)\.\(a\) Software Dependency Audit

\(b\) Sales Pivot Analysis

Mini’s skills raise stronger models’ average rewards\.On the 41 modified tasks, frozen Mini\-developed bundles raise Luna from 63\.9% to 74\.4% \(\+10\.5 pp\), with both conditions evaluated after freezing\. GPT\-5\.5 scores 71\.1% against its published original\-skill reference of 60\.5% \(\+10\.6 pp\)\. The two models improve on six shared tasks and on 20 distinct tasks in total \(Appendix[C](https://arxiv.org/html/2610.20926#A3)\)\.

Across all 86 tasks, Luna rises from 67\.3% to 71\.9%\. Modified bundles contribute \+5\.0 pp to this aggregate, while separately re\-executed unchanged bundles contribute \-0\.4 pp\. On the 41 modified tasks, Luna has 11 gains, 23 ties, and seven losses\.

Mini’s skills help stronger models even when Mini still makes errors\.Mini stays below 100\.0% on six of Luna’s 11 improved tasks\. Table[3](https://arxiv.org/html/2610.20926#S4.T3)shows how four executors satisfy different requirements with identical frozen bundles\. In Software Dependency Audit, original\-skill Mini and Luna scan beyond production dependencies\. With the revised skill, all final executions follow the intended scope\. Mini rises from 0\.0% to 66\.7% despite a remaining output\-field failure, while Luna,GPT\-5\.4, andGPT\-5\.5reach a mean task reward of 100\.0%\.

Sales Pivot Analysis reveals complementary difficulties \(Appendix[F\.1](https://arxiv.org/html/2610.20926#A6.SS1)\)\. Final Mini executions satisfy state\-value checks, but two omit Excel pivot objects; original\-skill Luna builds pivots but includes an excluded category\. The evolved skill specifies operation order and intermediate checks\. Luna adapts the abbreviated whitelist example to the input’s full state names and passes both state\-value and pivot\-object checks\. In one successful execution, Luna saves, reopens, and checks the source sheet before adding pivots\. A later call to the bundled helper fails because a dependency is missing, but the workbook already produced passes verification with full reward\.

Models can still misread input images or leave required output files unsaved\.In JPG OCR Statistics, Luna writes correctly structured workbooks but misreads receipt month tokens; the revised parsing rules leave the perception error unresolved\. In Paratransit Routing, two GPT\-5\.5 executions find and check better feasible solutions but fail to save them as the required output\. Appendix[F](https://arxiv.org/html/2610.20926#A6)provides these receiver execution records\.

## 5Conclusion

SkillNEOseparates the skill version used for deployment from the versions kept for further development, using comparative execution evidence to guide their revision\. The recorded searches show why this distinction matters\. Partially repaired bundles lead to later deployed descendants, and branches with lower overall reward contribute working procedures to other revisions\. On SkillsBench, development using GPT\-5\.4 Mini alone raises its average reward into GPT\-5\.5 and Luna’s original\-skill range\. The frozen bundles also benefit stronger executors on the same tasks\. These findings connect the value of an intermediate version to the further repairs and procedures it makes possible\. Future work should test the evolved skills on new instances within their intended task scope and study how to accumulate their improvements in a shared skill library\.

## References

- L\. A\. Agrawal, S\. Tan, D\. Soylu, N\. Ziems, R\. Khare, K\. Opsahl\-Ong, A\. Singhvi, H\. Shandilya, M\. J\. Ryan, M\. Jiang, C\. Potts, K\. Sen, A\. G\. Dimakis, I\. Stoica, D\. Klein, M\. Zaharia, and O\. KhattabGEPA: reflective prompt evolution can outperform reinforcement learning\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/0e9e708b6f48e14fd0ac29e167413f76-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p3.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Alzubiet al\.\(2026\)S\. Alzubi, N\. Provenzano, J\. Bingham, W\. Chen, and T\. VuEvoSkill: automated skill discovery for multi\-agent systems\.External Links:2603\.02766,[Link](https://arxiv.org/abs/2603.02766)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§2](https://arxiv.org/html/2610.20926#S2.p3.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Beyer and Schwefel \(2002\)H\. Beyer and H\. SchwefelEvolution strategies – a comprehensive introduction\.Natural Computing1\(1\),pp\. 3–52\(en\)\.External Links:[Document](https://dx.doi.org/10.1023/A%3A1015059928466),[Link](https://link.springer.com/article/10.1023/A:1015059928466)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p3.1),[§2](https://arxiv.org/html/2610.20926#S2.p3.1),[§3\.2](https://arxiv.org/html/2610.20926#S3.SS2.p2.2)\.
- Chang and Chen \(2026\)Y\. Chang and X\. ChenNaive prompt optimization: rethinking the need for complex prompt search\.External Links:2608\.27266,[Link](https://arxiv.org/abs/2608.27266)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Fernandoet al\.\(2024\)C\. Fernando, D\. S\. Banarse, H\. Michalewski, S\. Osindero, and T\. RocktäschelPromptbreeder: self\-referential self\-improvement via prompt evolution\.InProceedings of the 41st International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.235,pp\. 13481–13544\.External Links:[Link](https://proceedings.mlr.press/v235/fernando24a.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Gaoet al\.\(2026\)H\. Gao, H\. Chen, C\. Wang, S\. Guo, L\. Pang, Z\. Liu, H\. Shen, and X\. ChengSkillAudit: ground\-truth\-free skill evolution via paired trajectory auditing\.External Links:2606\.14239v1,[Link](https://arxiv.org/abs/2606.14239v1)Cited by:[§B\.3](https://arxiv.org/html/2610.20926#A2.SS3.p3.1),[§B\.3](https://arxiv.org/html/2610.20926#A2.SS3.p4.1),[Table 5](https://arxiv.org/html/2610.20926#A2.T5.12.2.1),[§1](https://arxiv.org/html/2610.20926#S1.p4.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Gaoet al\.\(2025\)Y\. Gao, Z\. Wang, J\. Chen, M\. Antkowiak, M\. Hu, J\. Kong, D\. Pratt, J\. Liu, E\. Ma, Z\. Hu, and E\. P\. XingscPilot: large language model reasoning toward automated single\-cell analysis and discovery\.InThe Thirty\-ninth Annual Conference on Neural Information Processing Systems,External Links:[Link](https://openreview.net/forum?id=Vzi96rTe4w)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Haoet al\.\(2024\)S\. Hao, Y\. Gu, H\. Luo, T\. Liu, X\. Shao, X\. Wang, S\. Xie, H\. Ma, A\. Samavedhi, Q\. Gao, Z\. Wang, and Z\. HuNew evaluation, library, and analysis of step\-by\-step reasoning with large language models\.InFirst Conference on Language Modeling,External Links:[Link](https://arxiv.org/abs/2404.05221v2)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Haoet al\.\(2023\)S\. Hao, Y\. Gu, H\. Ma, J\. J\. Hong, Z\. Wang, D\. Z\. Wang, and Z\. HuReasoning with language model is planning with world model\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 8154–8173\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.507),[Link](https://aclanthology.org/2023.emnlp-main.507/)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Huet al\.\(2026\)Z\. Hu, M\. Lu, Z\. Wang, J\. Ruan, C\. Chen, Z\. Pan, Y\. Guan, R\. Wang, Z\. Yu, C\. Zhang, and Y\. DingFlashEvolve: accelerating agent self\-evolution with asynchronous stage orchestration\.External Links:2605\.08520,[Link](https://arxiv.org/abs/2605.08520)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Jianget al\.\(2026\)Y\. Jiang, D\. Li, H\. Deng, B\. Ma, X\. Wang, Q\. Wang, and G\. YuSoK: agentic skills – beyond tool use in LLM agents\.External Links:2602\.20867,[Link](https://arxiv.org/abs/2602.20867)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Kang and Wen \(2026\)H\. Kang and M\. WenSkillLift: learning dense rubrics from sparse oracles for efficient skill evolution\.External Links:2609\.15396,[Link](https://arxiv.org/abs/2609.15396)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Liet al\.\(2026\)X\. Li, Y\. Liu, W\. Chen, B\. You, Z\. Di, Y\. He, S\. Zheng, K\. W\. Choe, J\. Sun, S\. Wang, C\. Tao, B\. Li, X\. Zhao, H\. Geng, X\. Wu, J\. Zhou, X\. Chen, H\. Xing, Y\. Li, Q\. Zeng, D\. Wang, Y\. Wang, R\. B\. Chaim, P\. Jiang, H\. Shen, L\. Kong, X\. Liu, R\. Wang, X\. Liu, J\. Li, X\. Lan, Y\. Lin, W\. Ye, J\. He, S\. Li, Y\. Zhang, Y\. Gao, Y\. Li, Z\. Ma, L\. Jing, T\. Wang, K\. Li, Y\. Xue, H\. Lyu, Y\. He, Y\. Tian, S\. Wu, B\. Wang, Y\. Gao, B\. Chen, L\. Liu, S\. Cheng, J\. Bao, S\. Tong, S\. Xu, T\. Y\. Zhuo, T\. Ye, Q\. Qi, M\. Li, L\. Liao, Z\. Tan, C\. Shi, X\. Tang, S\. Tankasala, B\. Yuan, Y\. Qian, J\. Tu, C\. Wang, Y\. Sun, W\. Wang, A\. Taylor, Z\. Yang, C\. Guan, Z\. Dong, X\. Zhang, S\. Dillmann, H\. Lee, and D\. SongSkillsBench: benchmarking how well agent skills work across diverse tasks\.External Links:2602\.12670,[Link](https://arxiv.org/abs/2602.12670)Cited by:[Appendix B](https://arxiv.org/html/2610.20926#A2.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§3\.1](https://arxiv.org/html/2610.20926#S3.SS1.p1.1),[§4\.1](https://arxiv.org/html/2610.20926#S4.SS1.p2.1)\.
- Linet al\.\(2026a\)H\. Lin, P\. Li, J\. Song, F\. Jiang, and T\. ZhangMUSE\-Autoskill: self\-evolving agents via skill creation, memory, management, and evaluation\.External Links:2605\.27366v2,[Link](https://arxiv.org/abs/2605.27366v2)Cited by:[Table 5](https://arxiv.org/html/2610.20926#A2.T5.12.3.1)\.
- Linet al\.\(2026b\)T\. Lin, C\. Sun, J\. Zhang, C\. Wei, H\. Yao, S\. Liu, X\. Zhang, L\. Liu, and J\. HuangOpenClaw\-Skill: collective skill tree search for agentic large language models\.External Links:2606\.16774,[Link](https://arxiv.org/abs/2606.16774)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Liuet al\.\(2026\)J\. Liu, M\. Hu, J\. Chen, J\. Kong, P\. Jagannatha, Y\. Gao, D\. Pratt, H\. Lee, Z\. Hu, T\. Ideker, W\. Wang, E\. P\. Xing, and Z\. WangHypoEvolve: genetic algorithms enable multi\-agent LLMs to discover scientific hypotheses\.External Links:2609\.15938,[Link](https://arxiv.org/abs/2609.15938)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Maet al\.\(2026\)E\. Ma, Y\. Zhou, W\. Huang, J\. Yang, H\. Ma, Z\. Wang, C\. Li, C\. Miao, P\. S\. Yu, and Z\. WangMemAudit: auditing long\-term agent memory via hidden user\-state recovery\.External Links:2606\.24595,[Link](https://arxiv.org/abs/2606.24595)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Niet al\.\(2026\)J\. Ni, Y\. Liu, X\. Liu, Y\. Sun, M\. Zhou, P\. Cheng, D\. Wang, E\. Zhao, X\. Jiang, and G\. JiangTrace2Skill: distill trajectory\-local lessons into transferable agent skills\.External Links:2603\.25158,[Link](https://arxiv.org/abs/2603.25158)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p4.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Olivetoet al\.\(2018\)P\. S\. Oliveto, T\. Paixão, J\. Pérez Heredia, D\. Sudholt, and B\. TrubenováHow to escape local optima in black box optimisation: when non\-elitism outperforms elitism\.Algorithmica80\(5\),pp\. 1604–1633\(en\)\.Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p3.1)\.
- OpenAI \(2026a\)OpenAICodex CLI\.Note:Software documentation\. Accessed September 17, 2026External Links:[Link](https://learn.chatgpt.com/docs/codex/cli)Cited by:[§4\.1](https://arxiv.org/html/2610.20926#S4.SS1.p3.1)\.
- OpenAI \(2026b\)OpenAIGPT\-5\.4 Mini model\.Note:API model documentation\. Accessed September 12, 2026External Links:[Link](https://developers.openai.com/api/docs/models/gpt-5.4-mini)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p5.1)\.
- Pryzantet al\.\(2023\)R\. Pryzant, D\. Iter, J\. Li, Y\. T\. Lee, C\. Zhu, and M\. ZengAutomatic prompt optimization with "gradient descent" and beam search\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,pp\. 7957–7968\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.494),[Link](https://aclanthology.org/2023.emnlp-main.494/)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Renet al\.\(2026\)Y\. Ren, H\. Zhang, L\. Xiao, J\. Huang, J\. Qiu, B\. Yu, Q\. Chen, and L\. LiuBranch2Skill: efficient skill evolution through reasoning trees\.External Links:2608\.08677,[Link](https://arxiv.org/abs/2608.08677)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Shastriet al\.\(2026\)Y\. P\. Shastri, A\. Eswaran, A\. Qidwai, P\. Thorat, and S\. JoshiBONSAI: evolvability\-guided tree search over skills\.External Links:2608\.07056,[Link](https://arxiv.org/abs/2608.07056)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§2](https://arxiv.org/html/2610.20926#S2.p4.1)\.
- Shinnet al\.\(2023\)N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. YaoReflexion: language agents with verbal reinforcement learning\.InThirty\-seventh Conference on Neural Information Processing Systems,External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Singlaet al\.\(2024\)S\. Singla, Z\. Wang, T\. Liu, A\. Ashfaq, Z\. Hu, and E\. P\. XingDynamic rewarding with prompt optimization enables tuning\-free self\-alignment of language models\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 21889–21909\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.1220/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.1220)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Sunet al\.\(2026\)R\. Sun, Z\. Zheng, Z\. Wang, and Z\. LuGraphSkillEvo: evolutionary optimization of graph\-structured agent skills\.External Links:2609\.21749,[Link](https://arxiv.org/abs/2609.21749)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§2](https://arxiv.org/html/2610.20926#S2.p3.1)\.
- Tanet al\.\(2026\)S\. Tan, L\. A\. Agrawal, R\. Sandadi, D\. Klein, K\. Sen, A\. G\. Dimakis, and M\. ZahariaAutomatically learning skills for coding agents\.Note:GEPA BlogExternal Links:[Link](https://gepa-ai.github.io/gepa/blog/2026/02/18/automatically-learning-skills-for-coding-agents/)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Wanget al\.\(2023\)G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. AnandkumarVoyager: an open\-ended embodied agent with large language models\.External Links:2305\.16291,[Link](https://arxiv.org/abs/2305.16291)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Wanget al\.\(2026a\)P\. Wang, R\. Cai, Z\. Wang, H\. Mei, Q\. Liu, P\. Li, and Z\. Wang∇\\nabla\-Reasoner: LLM reasoning via test\-time gradient descent in latent space\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=pEJAja73dk)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Wanget al\.\(2024\)X\. Wang, Z\. Wang, C\. Li, F\. Bai, H\. Luo, J\. Zhang, N\. Jojic, E\. Xing, and Z\. HuPromptAgent: strategic planning with large language models enables expert\-level prompt optimization\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/file/686a3f32067838c8dbb68da6e9e3cf69-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Wanget al\.\(2026b\)Z\. Wang, M\. Yan, J\. Bi, S\. Yan, V\. Tresp, and Y\. MaMetaSkill\-Evolve: recursive self\-improvement of LLM agents via two\-timescale meta\-skill evolution\.External Links:2607\.05297,[Link](https://arxiv.org/abs/2607.05297)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§2](https://arxiv.org/html/2610.20926#S2.p3.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Wanget al\.\(2026c\)Z\. Wang, F\. Bai, Z\. Luo, J\. Su, K\. Sun, X\. Yu, J\. Liu, K\. Zhou, C\. Cardie, M\. Dredze, Z\. Hu, and E\. P\. XingFIRE\-Bench: evaluating AI agents on the rediscovery of scientific insights\.InForty\-third International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=l3Yc9HS4US)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Wanget al\.\(2025\)Z\. Z\. Wang, J\. Mao, D\. Fried, and G\. NeubigAgent Workflow Memory\.InProceedings of the 42nd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.267,pp\. 63897–63911\.External Links:[Link](https://proceedings.mlr.press/v267/wang25bx.html)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Xiaet al\.\(2026\)P\. Xia, J\. Chen, H\. Wang, J\. Liu, K\. Zeng, Y\. Wang, S\. Han, Y\. Zhou, X\. Zhao, H\. Chen, Z\. Zheng, C\. Xie, and H\. YaoSkillRL: evolving agents via recursive skill\-augmented reinforcement learning\.External Links:[Link](https://arxiv.org/abs/2602.08234v2),2602\.08234v2Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Xionget al\.\(2026\)S\. Xiong, A\. Payani, O\. Gungordu, and F\. FekriDIVE: unlocking self\-improvement in frozen language models through diversity\-driven skill evolution\.External Links:2608\.12486,[Link](https://arxiv.org/abs/2608.12486)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p4.1)\.
- Yanget al\.\(2024\)C\. Yang, X\. Wang, Y\. Lu, H\. Liu, Q\. V\. Le, D\. Zhou, and X\. ChenLarge language models as optimizers\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2024/hash/3339f19c5fcee3ad74502947a32be9e6-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Yanget al\.\(2026\)Y\. Yang, Z\. Gong, W\. Huang, Q\. Yang, Z\. Zhou, Z\. Huang, Y\. Li, X\. Gao, Q\. Dai, B\. Liu, K\. Qiu, Y\. Yang, D\. Chen, X\. Yang, and C\. LuoSkillOpt: executive strategy for self\-evolving agent skills\.External Links:2605\.23904,[Link](https://arxiv.org/abs/2605.23904)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Yeet al\.\(2026\)H\. Ye, X\. He, V\. Arak, H\. Dong, and G\. SongMeta context engineering via agentic skill evolution\.InProceedings of the 43rd International Conference on Machine Learning,Proceedings of Machine Learning Research, Vol\.306,pp\. 148325–148368\.External Links:[Link](https://proceedings.mlr.press/v306/ye26g.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Yuet al\.\(2026a\)S\. Yu, G\. Li, W\. Shi, and P\. QiPolySkill: learning generalizable skills through polymorphic abstraction for continual learning\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/hash/e350897ed832f9cad6f2e223e7acad6d-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Yuet al\.\(2026b\)X\. Yu, F\. Bai, K\. Sun, H\. Miao, A\. Anand, Z\. Luo, K\. Zhou, and Z\. WangPrimeScientist: strategic allocation of research effort in autonomous research\.External Links:2609\.17846,[Link](https://arxiv.org/abs/2609.17846)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p2.1)\.
- Zhanget al\.\(2026a\)G\. Zhang, E\. Zhu, J\. Zhou, C\. Jia, and H\. WangSkillEvolver: skill learning as a meta\-skill\.External Links:2605\.10500,[Link](https://arxiv.org/abs/2605.10500)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhanget al\.\(2026b\)H\. Zhang, S\. Fan, H\. P\. Zou, Y\. Chen, Z\. Wang, J\. Zhou, C\. Li, W\. Huang, Y\. Yao, K\. Zheng, X\. \(\. Liu, X\. Li, and P\. S\. YuCoEvoSkills: self\-evolving agent skills via co\-evolutionary verification\.External Links:2604\.01687,[Link](https://arxiv.org/abs/2604.01687)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§1](https://arxiv.org/html/2610.20926#S1.p2.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhanget al\.\(2026c\)H\. Zhang, Q\. Long, J\. Bao, T\. Feng, W\. Zhang, H\. Yue, and W\. WangMemSkill: learning and evolving memory skills for self\-evolving agents\.External Links:2602\.02474,[Link](https://arxiv.org/abs/2602.02474)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhanget al\.\(2026d\)J\. Zhang, S\. Hu, C\. Lu, R\. Lange, and J\. CluneDarwin Gödel Machine: open\-ended evolution of self\-improving agents\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/hash/aa5f5e6eb6f613ec412f1d948dfa21a5-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p4.1)\.
- Zhanget al\.\(2026e\)Q\. Zhang, C\. Hu, S\. Upasani, B\. Ma, F\. Hong, V\. Kamanuru, J\. Rainton, C\. Wu, M\. Ji, H\. Li, U\. Thakker, J\. Zou, and K\. OlukotunAgentic context engineering: evolving contexts for self\-improving language models\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/hash/8a94ff6f922d995d7d3f4ebf4143e442-Abstract-Conference.html)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhaoet al\.\(2024\)A\. Zhao, D\. Huang, Q\. Xu, M\. Lin, Y\. Liu, and G\. HuangExpeL: LLM agents are experiential learners\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.38,pp\. 19632–19642\.External Links:[Link](https://ojs.aaai.org/index.php/AAAI/article/view/29936),[Document](https://dx.doi.org/10.1609/aaai.v38i17.29936)Cited by:[§2](https://arxiv.org/html/2610.20926#S2.p5.1)\.
- Zhenget al\.\(2025\)B\. Zheng, M\. Y\. Fatemi, X\. Jin, Z\. Z\. Wang, A\. Gandhi, Y\. Song, Y\. Gu, J\. Srinivasa, G\. Liu, G\. Neubig, and Y\. SuSkillWeaver: web agents can self\-improve by discovering and honing skills\.External Links:2504\.07079,[Link](https://arxiv.org/abs/2504.07079)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhouet al\.\(2026a\)H\. Zhou, S\. Guo, A\. Liu, Z\. Yu, Z\. Gong, B\. Zhao, Z\. Chen, M\. Zhang, Y\. Chen, J\. Li, R\. Yang, Q\. Liu, X\. Yu, J\. Zhou, N\. Wang, C\. Sun, and J\. WangMemento\-Skills: let agents design agents\.External Links:2603\.18743,[Link](https://arxiv.org/abs/2603.18743)Cited by:[§1](https://arxiv.org/html/2610.20926#S1.p1.1),[§2](https://arxiv.org/html/2610.20926#S2.p1.1)\.
- Zhouet al\.\(2026b\)Y\. Zhou, Z\. Dong, Z\. Wang, C\. Jin, S\. Zhao, B\. Guo, D\. Gu, L\. Zhang, M\. Zhou, and D\. N\. MetaxasEvidence over plans: online trajectory verification for skill distillation\.External Links:2605\.09192,[Link](https://arxiv.org/abs/2605.09192)Cited by:[Table 5](https://arxiv.org/html/2610.20926#A2.T5.12.4.1)\.

## Appendix ALimitations and Future Directions

Understanding skill search\.Our experiments evaluate the complete framework, and recorded histories characterize how its development paths are used\. Matched\-budget comparisons are needed to isolate the effects of non\-elitist selection, comparative execution evidence, and cross\-branch reuse\. Relevant selection policies include single\-path refinement, selection from the union of current parents and children, and a persistent historical top\-kkfrontier with the same number of parent slots\. Disabling archive revisitation would assess its contribution within each policy\. Such comparisons should use the same tasks, initial bundles, developer, verifier, model\-call budgets, and task\-execution budgets\. Search currently stops after repeated deployment holds; future work could assess stopping criteria that track progress across the development population\.

Generalization beyond development tasks\.We evaluate frozen bundles on the same tasks and input instances used during development\. Fresh executions measure performance on those instances but do not establish generalization to unseen tasks or inputs\. Future work should test whether these procedures remain effective on new task variants\.

Search and developer capability\.Mini performs all model\-based development roles in this study\. Extending the framework to developers with different capabilities and model families would help establish how search depth and developer capability jointly shape skill improvement\. A central question is whether and when deeper search can compensate for limited developer capability, and which capabilities remain necessary for productive skill development\.

Developing shared skill libraries\.SkillNEOevolves the bundle supplied for each task\. A shared library would need to accumulate improvements across tasks while managing interactions with previously learned procedures\. Future work could study how to combine broadly useful procedures with task\-specific guidance and assess revisions across the tasks that depend on them\.

## Appendix BDevelopment Protocol

Development starts from the skills supplied for each SkillsBench task\([Li et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib1)\), keeping theirSKILL\.mddocuments and optional tools and resources together as a bundle\. Harbor executes tasks;SkillNEOcompares traces, edits bundles, evaluates candidates, and selects versions\.

The planned per\-task configuration requests six children, promotes up to three, measures each promoted child and the incumbent over three repeats, and allows at most five generations\. A full generation with six admitted children and three promotions contains six screening executions, six additional promoted\-child executions, and three incumbent executions\. Search width counts candidate bundles and is distinct from execution concurrency\.

Algorithm 1SkillNEOuses non\-elitist replacement and comparative execution evidence\.1:Task

tt, frozen actor

AmA\_\{m\}, original

L0L\_\{0\};

b=6b=6,

k=3k=3,

R=3R=3,

G=5G=5
2:Deployment configuration

θ\\theta\(including

δoff,δaux\\delta\_\{\\rm off\},\\delta\_\{\\rm aux\}\), validation configuration

ϕ\\phi, archive flag

3:Evaluate root

00on

RRrepeats; initialize tree,

h0←0h\_\{0\}\\leftarrow 0,

B0←\{0\}B\_\{0\}\\leftarrow\\\{0\\\},

z←0z\\leftarrow 0,

gstop←0g\_\{\\rm stop\}\\leftarrow 0
4:ifroot is measured\-perfectthen return

L0L\_\{0\}
5:for

g=1,…,Gg=1,\\ldots,Gdo

6:Assemble root evidence \(

g=1g=1\) or previous\-population and saved execution evidence

7:Proposer specifies

bbedits from

Bg−1B\_\{g\-1\}; Skill Editor materializes candidate bundles

8:Apply Edit Validation

Ct,ϕC\_\{t,\\phi\}and permitted repairs; discard rejected drafts

9:Save distinct admitted survivors

DgD\_\{g\}and their edit edges

10:Screen

DgD\_\{g\}once; rank complete children by single\-execution reward

11:iffewer than two children have complete screening recordsthen

12:Interrupt with

hg−1,Bg−1,zh\_\{g\-1\},B\_\{g\-1\},zunchanged; resume generation

ggafter completion

13:endif

14:Promote up to

kktop\-ranked screened children

15:Add repeats

2:R2\{:\}Rto promoted children; measure

hg−1h\_\{g\-1\}on

RRfresh repeats

16:Check task, model, environment, and verifier consistency

17:Rank complete

RR\-execution means, including screening for promoted children

18:ifthe incumbent has no complete comparison recordthen

19:

hg←h\_\{g\}\\leftarrowhighest\-ranked complete bundle

20:else

21:

c←c\\leftarrowhighest\-ranked child;

h←hg−1h\\leftarrow h\_\{g\-1\}
22:

d←Sg​\(c\)−Sg​\(h\)d\\leftarrow S\_\{g\}\(c\)\-S\_\{g\}\(h\);

a←Ag​\(c\)−Ag​\(h\)a\\leftarrow A\_\{g\}\(c\)\-A\_\{g\}\(h\)
23:

w←\[d\>δoff\]∨\[d=0∧a\>δaux\]w\\leftarrow\[d\>\\delta\_\{\\rm off\}\]\\lor\[d=0\\ \\land\\ a\>\\delta\_\{\\rm aux\}\]
24:

hg←ch\_\{g\}\\leftarrow cif

wwand

Acceptθ⁡\(c,h,𝐐g,E≤g\)=1\\operatorname\{Accept\}\_\{\\theta\}\(c,h;\\mathbf\{Q\}\_\{g\},E\_\{\\leq g\}\)=1; otherwise

hg←hh\_\{g\}\\leftarrow h
25:endif

26:

gstop←gg\_\{\\rm stop\}\\leftarrow g
27:

z←0z\\leftarrow 0if a new child is deployed; otherwise

z←z\+1z\\leftarrow z\+1
28:

Bg←BgordB\_\{g\}\\leftarrow B\_\{g\}^\{\\mathrm\{ord\}\}from Equation[4](https://arxiv.org/html/2610.20926#S3.E4)

29:On an enabled negativeHold, use eligible archive survivors with

RRrepeats first, then thinner records

30:If archive revisit finds no survivor, retain the ordinary set or stop if it is also empty

31:Append measured root if absent and

\|Bg\|<2\|B\_\{g\}\|<2
32:if

z≥2z\\geq 2or

\(z≥1∧100−Sg​\(hg\)<δoff\)\(z\\geq 1\\ \\land\\ 100\-S\_\{g\}\(h\_\{g\}\)<\\delta\_\{\\rm off\}\)then break

33:endfor

34:return

LhgstopL\_\{h\_\{g\_\{\\rm stop\}\}\}⊳\\trianglerightDeployment incumbent

### B\.1Search configuration

With generation limitGG, widthbb, promotion countkk, and repeat countRR, the complete planned schedule containsR\+G⁡\[b\+k⁡\(R−1\)\+R\]R\+G\[b\+k\(R\-1\)\+R\]task executions, giving 78 for\(G,b,k,R\)=\(5,6,3,3\)\(G,b,k,R\)=\(5,6,3,3\)\. Actual execution counts depend on the completed generations and measured candidates; proposal and skill\-editing calls and final evaluations are separate\. Algorithm[1](https://arxiv.org/html/2610.20926#alg1)summarizes the loop\.

The later\-generation proposer chooses its primary parent from the eligible pool, with a per\-parent cap of⌈b/\|Bg−1\|⌉\+2\\lceil b/\|B\_\{g\-1\}\|\\rceil\+2\. Graft donors are drawn from the same eligible pool, while execution evidence used for diagnostic comparisons also includes screening\-only candidates\. Diagnostic summaries include up to four bundle versions from each end of the ranking\. The later\-generation prompt lists edit directions previously measured as net losses and asks the Proposer to avoid closely similar directions with the same operator and target skill\. This proposal\-level constraint leaves parent eligibility to Equation[4](https://arxiv.org/html/2610.20926#S3.E4)and the archive rules below\.

Operators include rewriting, extending, pruning, removing, rerouting, adding executable support, and grafting from an eligible donor\. Every child retains one primary parent edge, including grafted children\. Output checks and deduplication can reduce the requested six children to fewer distinct bundles\. Under configurationϕ\\phi, the edit\-admission check isCt,ϕ​\(Lp⁡\(v\),Lv,av\)C\_\{t,\\phi\}\(L\_\{p\(v\)\},L\_\{v\},a\_\{v\}\); its constraints and repair paths appear in Appendix[B\.2](https://arxiv.org/html/2610.20926#A2.SS2)\.

Execution feedback\.The Proposer combines a fixed prompt template with agent\-visible traces, rewards, and filtered verifier findings\. Feedback filtering and candidate\-content checks address evaluator\-private answer material; public task requirements and observed behavior guide revisions\. Appendix[B\.2](https://arxiv.org/html/2610.20926#A2.SS2)specifies these checks\.

Screening and repeated evaluation\.For bundleLvL\_\{v\}, the mean reward over a nonempty execution batchQQwith outcomesyv,qy\_\{v,q\}is

J^v​\(Q\)=1\|Q\|​∑q∈QVt​\(yv,q\)\.\\widehat\{J\}\_\{v\}\(Q\)=\\frac\{1\}\{\|Q\|\}\\sum\_\{q\\in Q\}V\_\{t\}\(y\_\{v,q\}\)\.\(6\)The execution batchQQidentifies the runs used in each comparison\. The root is measured on three repeats and becomes the first incumbent; search stops if all three receive perfect reward\. Otherwise, every admitted child receives one screening execution\. Infrastructure failures are separated from official timeouts, memory limits, and verifier failures\. Incomplete records do not enter the ranking\. Screening ranks children by their single\-execution official reward, using compatible auxiliary scores to break ties\. Promotion requires at least two children with complete screening records\. If fewer are available, execution of the generation is interrupted without a deployment decision or a hold\-count update\. The saved incumbent and parent set are retained, and the same generation can resume after the missing screening measurements are completed\. Up to three top\-ranked complete children retain their screening measurement and receive repeats two and three; their resulting three\-execution means determine the deployment ranking\. The incumbent receives three fresh repeats interleaved in the same comparison batch without using a promotion slot\. Deployment compares only fully replicated records with matching task, model, environment, and verifier settings\.

Deployment acceptance\.For the best fully measured childcgc\_\{g\}, deployment updates as

hg=\{cg,Acceptθ⁡\(cg,hg−1,𝐐g,E≤g\)=1,hg−1,otherwise \(Hold\),h\_\{g\}=\\begin\{cases\}c\_\{g\},&\\operatorname\{Accept\}\_\{\\theta\}\(c\_\{g\},h\_\{g\-1\};\\mathbf\{Q\}\_\{g\},E\_\{\\leq g\}\)=1,\\\\ h\_\{g\-1\},&\\text\{otherwise \(\{Hold\}\)\},\\end\{cases\}\(7\)where𝐐g=\(Qg,v\)v\\mathbf\{Q\}\_\{g\}=\(Q\_\{g,v\}\)\_\{v\}records the execution batch for each compared node andE≤gE\_\{\\leq g\}collects saved development evidence\. The deployment configurationθ\\thetaspecifies the improvement margins and historical\-score checks\. These evidence arguments are suppressed below\.

WriteSg​\(v\)S\_\{g\}\(v\)for a bundle’s official mean reward in the current comparison batch andAg​\(v\)A\_\{g\}\(v\)for its auxiliary score, both on a scale from 0 to 100\. Candidates rank first bySgS\_\{g\}, then byAgA\_\{g\}\. The auxiliary score uses the percentage of passed verifier tests only when the compared evaluations have compatible test coverage; otherwise it uses official reward\. The incumbent wins an exact tie on both scores; candidate ties use a deterministic tie\-break\.

Letccbe the highest\-ranked fully measured child andhhthe incumbent\. For the current decision, writeS​\(v\)=Sg​\(v\)S\(v\)=S\_\{g\}\(v\),A​\(v\)=Ag​\(v\)A\(v\)=A\_\{g\}\(v\),Δ​S=S⁡\(c\)−S⁡\(h\)\\Delta S=S\(c\)\-S\(h\), andΔ​A=A⁡\(c\)−A⁡\(h\)\\Delta A=A\(c\)\-A\(h\)\. For the strict\-floor rule, letPg​\(v\)P\_\{g\}\(v\)be the mean official reward on the 0–100 scale over saved executions of bundlevvin this task search, up to and including the current comparison batch\. This pool includes its child evaluations, subsequent incumbent re\-evaluations, and root initialization where applicable\. Records sharing an execution identifier are first averaged, so each distinct execution contributes once\. At each deployment decision, recomputeFgF\_\{g\}as the maximum availablePg​\(v\)P\_\{g\}\(v\)forv∈\{0,h\}v\\in\\\{0,h\\\}; if neither pooled score is available, setFg=−∞F\_\{g\}=\-\\infty\. The strict\-floor acceptance rule uses improvement marginsδoff\\delta\_\{\\rm off\}andδaux\\delta\_\{\\rm aux\}for official and auxiliary scores and accepts a child according to

Acceptθ\(c,h\)=\[S\(c\)\>Fg∧\(ΔS\>δoff∨\(ΔS=0∧ΔA\>δaux\)\)\]\.\\operatorname\{Accept\}\_\{\\theta\}\(c,h\)=\\mathbf\{1\}\\\!\\left\[\{S\(c\)\>F\_\{g\}\}\\;\\land\\;\\bigl\(\\Delta S\>\\delta\_\{\\rm off\}\\;\\lor\\;\(\\Delta S=0\\;\\land\\;\\Delta A\>\\delta\_\{\\rm aux\}\)\\bigr\)\\right\]\.\(8\)The improvement margins are configurable, with defaultsδoff=0\\delta\_\{\\rm off\}=0andδaux=30\\delta\_\{\\rm aux\}=30pp\. The auxiliary margin limits incumbent replacement due to small auxiliary\-score fluctuations when official scores tie\. Without a complete incumbent batch, selection uses the highest\-ranked fully measured bundle\.

Developmental selection and archive revisit\.Ordinary parent selection takes the topkkfully measured current\-generation children whose complete bundles remain available, excluding the incumbent’s repeat\-evaluation entries\. The previous parent pool does not enter this ranking, so its best member need not remain eligible\. Neither parent\-relative improvement nor acceptance for deployment is required\. Rejected drafts and deduplicated bundles cannot become parents\. If fewer than two parents remain after selection, the measured root is added if absent\.

When enabled, archive revisit activates on aHoldwithΔ​S<0\\Delta S<0\. It retires fully measured children that score below the incumbent\. It also retires a parent if that parent has an available score in the same comparison and all of its evaluated children score below that parent\. Archive candidates are unretired nodes with no children remaining after deduplication and retirement\. Eligible survivors withRRrepeats are ranked first by official score\. If that tier is empty, the remaining measured survivors form the fallback tier\. If no archive survivor exists, selection uses the ordinary parent set when available; otherwise search stops as exhausted\. The measured root is then added if fewer than two parents remain and it is absent\. This revisit occurs in nine decisions across seven tasks\. Seven decisions use fully repeated survivors, and two use one\-repeat fallback candidates\.

Stopping\.Search returns the root if its initial measurements are perfect\. Deployment updates reset the hold counter; holds increment it\. Search stops after two consecutive holds, after a hold with100−Sg​\(hg\)<δoff100\-S\_\{g\}\(h\_\{g\}\)<\\delta\_\{\\rm off\}, at the generation limit, or on archive exhaustion\.

### B\.2Candidate construction and checks

Edit Validation programmatically checks operation constraints, referenced files, new scripts, and introduced content\. Fixable failures return to the Skill Editor; evaluation uses only admitted bundles\.

The Proposer receives public instructions, agent\-visible execution evidence, official rewards, and filtered evaluator findings\. The Skill Editor receives the edit specification and current skill\. Public task requirements and agent\-computed results can guide task\-specific changes; the checks target evaluator\-private answers copied into deployed artifacts\.

Before proposal generation, verifier\-text filtering removes comparison\-shaped answer text and known private literals\. Redaction and rejection use reference material from test and specification files, excluding the reference solution\. Where applicable, matching rules exclude public\-instruction vocabulary and imported API names\.

After editing, newSKILL\.mdlines and added or replaced support files are checked for exact private five\-word sequences, normalized lines, literals, and distinctive identifiers\. When a match triggers rejection, the Skill Editor receives a generic message without the matched answers\. Fixable rejections permit one revision; a missing promised executable permits an additional repair limited to that file\. Repaired artifacts are checked again, and rejected children do not enter evaluation\.

A configurable content heuristic uses default rejection thresholds of 20 new bare numeric entries, or six new entries on lines containing values with at least four decimal places\. Validation also checks for instructions that depend on unavailable grader messages and missing promised helper files\. Script checks execute Python helpers and check syntax for other supported languages\. Text\-only revisions do not require a helper check\. Broader grounding or skill\-routing concerns may be reported without rejecting the candidate\.

These checks target newly introduced content and specified matching patterns\. Admitted bundles are then executed and scored by the task verifier\.

Table 4:No\-skill and original\-skill references\. Original skills increase average reward in each listed configuration on the aligned 86 tasks\. Rewards are percentages andΔ\\Deltais in percentage points\. Official rows are reaggregated from the published per\-task registry\. Differences are computed from the displayed means\. No Luna no\-skill result has been measured\.
### B\.3Evaluation settings and comparison context

Final evaluation uses the frozen task\-bundle assignment and the model settings in Section[4\.1](https://arxiv.org/html/2610.20926#S4.SS1)\.

Reference source\.The official rows in Table[4](https://arxiv.org/html/2610.20926#A2.T4)and the GPT\-5\.5 original\-skill reference use the[published SkillsBench results](https://github.com/benchflow-ai/skillsbench-trajectories/blob/cb640f73cdea8c67422cae3cc37b913a7c080871/website-data/results-registry.json), reaggregated over the 86 tasks\. GPT\-5\.4’s 51\.7% original\-skill reference comes from[Vals](https://www.vals.ai/benchmarks/skillsbench)and uses 87 tasks, OpenHands, and xhigh reasoning\. The \+13\.8 pp difference between our 86\-task Codex final score and this reference compares different settings\.

Table 5:Reported skill\-development results on SkillsBench\. Average rewards \(%\) use each study’s task cohort and evaluation protocol\.Humandenotes each study’s human\-skill reference;Developeduses its resulting skills\. SkillEvolver and CoEvoSkills instead report complete\-pass frequencies\. SkillAudit, MUSE\-Autoskill, andSkillNEOuse the listed model for development and execution; SPARK uses stronger\-model teacher trajectories; SPARK’s scores follow Section 4\.1 of the source\. SkillAudit andSkillNEOrefine supplied skills; MUSE\-Autoskill and SPARK generate skills, with Human as a separate reference\. MUSE counts its 28 uncovered tasks as zero\. SPARK describes an 86\-task suite; exact Mini task alignment is unavailable\. ForSkillNEO, 41 modified bundles receive new final executions; 45 retained bundles share their original Mini measurements\.StudyExecutorTasksHumanDevelopedSkillAudit\([Gao et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib3)\)Opus 4\.88956\.773\.9MUSE\-Autoskill\([Lin et al\., 2026a](https://arxiv.org/html/2610.20926#bib.bib29)\)GPT\-5\.57559\.6753\.42SPARK\([Zhou et al\., 2026b](https://arxiv.org/html/2610.20926#bib.bib30)\)GPT\-5\.4 Mini86≈47\\approx 47≈52\\approx 52SkillNEO\(this study\)GPT\-5\.4 Mini8649\.166\.4Reported skill\-development results on SkillsBench\.Table[5](https://arxiv.org/html/2610.20926#A2.T5)summarizes average reward reported by three prior studies alongside our results\. Reported averages are 66\.4% forSkillNEO, 53\.42% for MUSE\-Autoskill with GPT\-5\.5, and approximately 52% for Mini with SPARK skills, each under its own task set and evaluation protocol\. SkillAudit reports 73\.9%, using Opus 4\.8 for task execution, PACE auditing, skill editing, and final evaluation\([Gao et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib3)\)\.

Comparative auditing and continued development\.The studies also differ in how they develop skills\. MUSE\-Autoskill creates skills from executions, and SPARK distills stronger\-model trajectories\. To repair supplied skills, SkillAudit\([Gao et al\., 2026](https://arxiv.org/html/2610.20926#bib.bib3)\)uses PACE to compare executions with and without the current skill, combines these diagnostics with a fixed structuralverifier, and can roll back to the previous saved skill version\. Its reported loop maintains one active skill state\.SkillNEOalso develops supplied bundles with Mini and permits continued editing of children without parent\-relative improvement\. Parent\-child and sibling records organize evidence across revisions, and grafting allows procedures from another saved branch to enter a candidate\. These mechanisms preserve alternatives both as starting points for further edits and as sources of useful procedures; Section[4\.3](https://arxiv.org/html/2610.20926#S4.SS3)traces both uses in the recorded executions\. SkillAudit excludes external task rewards during evolution, whereasSkillNEOuses benchmark rewards to evaluate candidates and select versions\.

## Appendix CAnalysis Across Receiving Models

We analyze the three\-repeat task means from the 86\-task evaluation\. The 41 modified tasks divide into 29 Mini\-positive and 12 Mini\-tied tasks under the±0\.05\\pm 0\.05pp tie band; the other 45 tasks retain their original bundles\. These Mini\-outcome subgroups are defined by the final evaluation results\.

Table 6:Receiver rewards grouped by Mini’s outcome\. GPT\-5\.5 and Luna have larger mean gains on tasks where Mini improves than on tasks where Mini ties\. Scores report average reward \(%\);Gapis evolved minus original reward in percentage points\. The 41 modified tasks are grouped post hoc by Mini’s final evaluation outcomes, using its±0\.05\\pm 0\.05pp display band\. Evolved denotes the same frozen Mini\-developed bundle for each task\. GPT\-5\.4 reports frozen\-skill endpoints; original\-skill subgroup scores are not reported\. Evaluation settings follow Section[4\.1](https://arxiv.org/html/2610.20926#S4.SS1)\.Reward by Mini outcome\.Table[6](https://arxiv.org/html/2610.20926#A3.T6)reports receiver rewards and their differences from available original\-skill references\. In the Mini\-improved group, GPT\-5\.5 and Luna gain 13\.8 and 18\.3 pp, respectively\. In the Mini\-tied group, GPT\-5\.5 gains 2\.8 pp while Luna loses 8\.3 pp\.

Why selected revisions can tie the original bundle in final evaluation\.Selection compares a candidate with that generation’s incumbent, whereas final evaluation compares the frozen bundle with the original skill in separate execution batches\. Both the reference bundle and observed reward can therefore differ\. Eleven of the 12 Mini\-tied bundles were adopted with higher official reward than the incumbent\. Grid was adopted at 100 with a 33\.3\-point incumbent gap, while its final original and evolved scores were both 66\.7\.

The final ties consist of four ceiling cases, three partial\-score cases, and five zero\-score cases\. Perfect original reward leaves no room for observed improvement; when both conditions yield zero reward, neither resolves the recorded failures\.

Absolute success and agreement\.Among the 41 modified tasks, 14 achieve 100\.0% average reward with all three receivers, 15 with one or two, and 12 with none\. Achieving 100\.0% requires all three saved repeats to receive full reward\. Three tasks receive zero with all three receivers: Manufacturing Codebook Normalization, Paratransit Routing, and Quantum Numerical Simulation\. Comparisons with original skills distinguish full reward from improvement for GPT\-5\.5 and Luna\.

Table 7:Joint directions of receiver reward changes\. Gains and losses need not coincide across GPT\-5\.5 and Luna on the 41 modified tasks\. Rows use GPT\-5\.5 reference differences; columns use Luna differences from post\-freeze evaluations\. Both axes use the existing±0\.05\\pm 0\.05pp display band\. Original\-skill reference sources appear in Appendix[B](https://arxiv.org/html/2610.20926#A2)\.Direction depends on the receiving model and its comparison\.Table[7](https://arxiv.org/html/2610.20926#A3.T7)groups tasks by the directions of their GPT\-5\.5 and Luna reward differences\. Six tasks gain in both, while 20 gain in at least one\. Four have opposite nonzero signs\. Civ6 Adjacency Optimizer, Grid Dispatch, and Organize Messy Files decrease in both; Appendix[F](https://arxiv.org/html/2610.20926#A6)details each model’s failures\.

Table 8:Contributions to Luna’s overall reward difference\. Gains on modified tasks outweigh the decrease on retained tasks\. Rewards are percentages;Δ\\DeltaandFull\-set contributionare in percentage points\. The Mini\-positive and Mini\-tied subsets partition the 41 modified tasks\. The 41 modified and 45 retained tasks partition the benchmark\. Their contributions aren​Δ/86n\\Delta/86and sum using unrounded values\. Mini\-positive/tied strata are defined post hoc from Mini’s final evaluation outcomes\.Contributions to the full\-cohort differences\.Table[8](https://arxiv.org/html/2610.20926#A3.T8)weights each subset’s mean reward difference by its share of the 86 tasks\. Luna’s modified and retained subsets contribute approximately\+4\.99\+4\.99and−0\.39\-0\.39pp to the\+4\.60\+4\.60pp aggregate gain; the retained subset’s own mean differs by−0\.74\-0\.74pp between separate original\- and final\-assignment execution batches, although its skill bundles are identical\. GPT\-5\.5’s contributions of\+5\.05\+5\.05and−1\.55\-1\.55pp yield a\+3\.50\+3\.50pp difference\.

Table 9:Luna transfer by observed original performance\. Among the 41 modified tasks, the group with zero original reward has the largest mean gain\. Rewards are percentages,Δ\\Deltais in percentage points, andG/T/Lcounts gains, ties, and losses with the±0\.05\\pm 0\.05pp display tolerance of Table[1](https://arxiv.org/html/2610.20926#S4.T1)\. Groups are defined post hoc from Luna’s three original\-skill repeats\.To relate reuse gains to starting performance, Table[9](https://arxiv.org/html/2610.20926#A3.T9)groups the 41 modified tasks by their observed reward on Luna when using the original skills\. Four of the eight tasks with zero original reward improve and four tie\. Outcomes are more mixed among the 16 tasks with partial original reward: seven improve, three tie, and six lose reward\. Of the 17 tasks with perfect original reward, 16 retain perfect reward and one receives a lower reward\.

## Appendix DDomain Results

Mini improves in every domain, while Luna’s domain\-level differences include gains and losses\. Table[2](https://arxiv.org/html/2610.20926#S4.T2)reports Mini’s domain scores from Figure[3](https://arxiv.org/html/2610.20926#S4.F3)\. Table[10](https://arxiv.org/html/2610.20926#A4.T10)adds Mini and Luna differences on the modified\-task subsets and Luna’s full\-domain differences\.

Table 10:Domain\-level improvement and transfer\. Mini improves in every domain, while Luna differences vary by domain\. Differences are percentage points\.AllandModifiedgive task counts; the modified\-task and full\-domain differences use their respective denominators\.### D\.1Resolved and remaining workflow obstacles

To contrast resolved and remaining workflow obstacles, we examine original and final skills with three Mini executions per condition for four tasks \(24 executions\)\.

Removing an execution barrier\.In SimPO Code Reproduction, original executions log Python 3\.12, violating the required version 3\.10 and stopping verification at the Python\-version assertion\. Comparing the saved outputs separately shows that all original runs’ 100 loss values numerically match those from passing final executions\. The evolved skill makes creating and activating the required environment precede installation, logging, and testing\. All final executions use Python 3\.10 and pass, raising reward from 0\.0 to 100\.0\.

Data\-to\-D3 also rises from 0\.0 to 100\.0\. Original executions write the required files and check syntax, but do not perform browser or DOM interaction checks; all fail tooltip behavior\. The evolved skill explicitly requires checking the rendered page and linked interactions\. All final executions perform interaction checks \(two with Chromium and one with jsdom\) and pass all 15 tests\. One check reveals that repeated selection clears the selected row; the agent changes toggling to idempotent selection before submission\. The skill guides the agent’s construction of execution\-time checks\.

Failures can remain after a targeted edit\.Adaptive Cruise Control rises from 33\.3 to 66\.7\. Its evolved skill requires checking fresh simulation outputs, but names a helper path that is unavailable during execution\. All final runs implement their own checks\. The remaining failed run checks aggregate metrics but leaves emergency behavior absent, failing safety and scenario checks\. The bundled checker omits these conditions and cannot verify the complete controller\.

Manufacturing Codebook Normalization exposes inconsistent rule application, with a final reward of 0\.0\. The revised rule requires component evidence in the raw segment for certain electrical\-fault labels\. All final runs read the skill, but the two completed runs also admit contextual test\-item evidence; one further replaces the required raw\-text cue with a higher matching threshold\. Both retain mismatched component assignments and pass 15 of 16 tests, failing the remaining codebook check based on lexical overlap\. Another final run times out without the required output\.

## Appendix EComplete Task\-Level Results

Table 11:Complete task\-level rewards\. Each row reports mean reward \(%\) on one of the 86 tasks\. M denotes modified bundles; R denotes byte\-identical retained bundles\. O/E denote original/final assignments\. Mini’s R rows reuse the same original three\-run score in both columns\. Evaluation follows Section[4\.1](https://arxiv.org/html/2610.20926#S4.SS1); Appendix[B](https://arxiv.org/html/2610.20926#A2)identifies the sources of the original\-skill references\.TaskMini OMini E5\.5 O5\.5 ELuna OLuna E3d\-scan\-calcR100\.0100\.0100\.0100\.0100\.0100\.0ada\-bathroom\-plan\-repairR33\.333\.3100\.0100\.0100\.0100\.0adaptive\-cruise\-controlM33\.366\.733\.3100\.066\.7100\.0azure\-bgp\-oscillation\-route\-leakR0\.00\.00\.00\.00\.00\.0citation\-checkR100\.0100\.0100\.0100\.0100\.0100\.0civ6\-adjacency\-optimizerM0\.023\.395\.063\.363\.331\.7court\-form\-fillingM33\.366\.733\.366\.733\.3100\.0crystallographic\-wyckoff\-position\-analysisM75\.7100\.0100\.0100\.0100\.0100\.0dapt\-intrusion\-detectionR100\.0100\.0100\.0100\.0100\.0100\.0data\-to\-d3M0\.0100\.033\.3100\.00\.00\.0debug\-trl\-grpoM41\.761\.791\.783\.3100\.0100\.0dialogue\-parserM83\.383\.383\.383\.3100\.0100\.0drone\-planning\-controlM42\.254\.456\.770\.084\.467\.8dynamic\-object\-aware\-egomotionR0\.00\.033\.30\.066\.70\.0earthquake\-phase\-associationR66\.766\.7100\.0100\.0100\.0100\.0earthquake\-plate\-calculationR100\.0100\.0100\.0100\.0100\.0100\.0econ\-detrending\-correlationM0\.0100\.0100\.0100\.0100\.0100\.0edit\-pdfM0\.033\.3100\.0100\.0100\.0100\.0energy\-ac\-optimal\-power\-flowR100\.0100\.0100\.0100\.0100\.0100\.0energy\-market\-pricingM66\.7100\.033\.366\.766\.70\.0energy\-unit\-commitmentR0\.00\.0100\.00\.00\.00\.0enterprise\-information\-searchR0\.00\.00\.00\.00\.00\.0exam\-block\-sequencingR100\.0100\.0100\.066\.70\.00\.0exceltable\-in\-pptR100\.0100\.0100\.0100\.0100\.0100\.0exoplanet\-detection\-periodM0\.066\.7100\.0100\.066\.7100\.0financial\-modeling\-qaR66\.766\.7100\.0100\.0100\.033\.3fix\-build\-agentopsM66\.7100\.066\.766\.7100\.0100\.0fix\-build\-google\-autoR0\.00\.00\.066\.766\.733\.3fix\-druid\-loophole\-cveR100\.0100\.066\.766\.7100\.0100\.0fix\-erlang\-ssh\-cveM100\.0100\.033\.366\.7100\.0100\.0fix\-visual\-stabilityR100\.0100\.033\.30\.066\.7100\.0flink\-queryR0\.00\.033\.30\.00\.033\.3flood\-risk\-analysisR0\.00\.00\.00\.033\.30\.0glm\-lake\-mendotaR100\.0100\.00\.0100\.0100\.0100\.0gravitational\-wave\-detectionM0\.0100\.0100\.0100\.0100\.0100\.0grid\-dispatch\-operatorM66\.766\.766\.733\.366\.70\.0hvac\-controlM100\.0100\.0100\.0100\.033\.3100\.0invoice\-fraud\-detectionM0\.0100\.00\.0100\.00\.0100\.0jax\-computing\-basicsR100\.0100\.0100\.0100\.0100\.0100\.0jpg\-ocr\-statM0\.066\.733\.333\.333\.30\.0lab\-unit\-harmonizationR100\.0100\.0100\.0100\.0100\.0100\.0lake\-warming\-attributionM66\.7100\.066\.733\.366\.7100\.0latex\-formula\-extractionR33\.333\.30\.00\.0100\.00\.0lean4\-proofR100\.0100\.0100\.0100\.0100\.0100\.0llm\-prefix\-cache\-replayR100\.0100\.0100\.0100\.0100\.0100\.0manufacturing\-codebook\-normalizationM0\.00\.00\.00\.00\.00\.0manufacturing\-equipment\-maintenanceR33\.333\.3100\.066\.766\.7100\.0manufacturing\-fjsp\-optimizationR0\.00\.033\.30\.0100\.066\.7mario\-coin\-countingM100\.0100\.00\.0100\.0100\.0100\.0mars\-clouds\-clusteringM33\.366\.7100\.0100\.0100\.0100\.0multilingual\-video\-dubbingM33\.3100\.033\.3100\.0100\.0100\.0offer\-letter\-generatorR100\.0100\.0100\.0100\.0100\.0100\.0organize\-messy\-filesM0\.00\.0100\.033\.366\.70\.0paper\-anonymizerM0\.033\.3100\.0100\.0100\.0100\.0parallel\-tfidf\-searchM33\.3100\.0100\.0100\.066\.7100\.0paratransit\-routingM0\.033\.3100\.00\.00\.00\.0pddl\-airport\-planningM33\.3100\.066\.7100\.0100\.0100\.0pddl\-tpp\-planningR100\.0100\.0100\.0100\.033\.366\.7pdf\-excel\-diffR100\.0100\.0100\.0100\.0100\.0100\.0powerlifting\-coef\-calcR100\.0100\.0100\.0100\.0100\.0100\.0pptx\-reference\-formattingR100\.0100\.0100\.0100\.0100\.0100\.0protein\-expression\-analysisR100\.0100\.0100\.0100\.033\.3100\.0python\-scala\-translationM0\.0100\.033\.3100\.00\.0100\.0quantum\-numerical\-simulationM0\.00\.00\.00\.00\.00\.0r2r\-mpc\-controlM100\.0100\.066\.7100\.0100\.066\.7radar\-vital\-signsR66\.766\.7100\.0100\.066\.7100\.0react\-performance\-debuggingM66\.766\.766\.733\.366\.766\.7reserves\-at\-risk\-calcR0\.00\.00\.00\.00\.00\.0sales\-pivot\-analysisM0\.033\.30\.00\.00\.0100\.0sec\-financial\-reportM33\.366\.766\.70\.066\.766\.7seismic\-phase\-pickingR33\.333\.3100\.0100\.0100\.0100\.0setup\-fuzzing\-pyM0\.022\.055\.383\.05\.383\.0shock\-analysis\-demandR0\.00\.00\.00\.00\.00\.0shock\-analysis\-supplyM0\.00\.00\.00\.066\.766\.7simpo\-code\-reproductionM0\.0100\.00\.0100\.0100\.0100\.0software\-dependency\-auditM0\.066\.766\.7100\.00\.0100\.0spring\-boot\-jakarta\-migrationR100\.0100\.0100\.0100\.0100\.0100\.0suricata\-custom\-exfilM0\.00\.0100\.0100\.0100\.0100\.0syzkaller\-ppdev\-syzlangR100\.0100\.0100\.0100\.0100\.0100\.0threejs\-structure\-parserR100\.0100\.0100\.0100\.033\.3100\.0threejs\-to\-objR100\.0100\.0100\.0100\.0100\.0100\.0tictoc\-unnecessary\-abort\-detectionM81\.7100\.0100\.0100\.0100\.0100\.0travel\-planningR100\.0100\.0100\.0100\.0100\.0100\.0video\-silence\-removerR0\.00\.00\.00\.00\.00\.0weighted\-gdp\-calcR100\.0100\.0100\.0100\.0100\.0100\.0xlsx\-recover\-dataR0\.00\.00\.00\.00\.00\.0Table[11](https://arxiv.org/html/2610.20926#A5.T11)reports the full 86\-task comparison underlying the aggregate results\. Each row corresponds to one task and its complete collection of skills\. A modified marker denotes an evolved bundle; a retained marker denotes a byte\-identical original bundle\. Mini’s 41 modified bundles use post\-evolution final executions\. Its 45 retained rows represent the same skill configuration in both columns and use the same original three\-run score\. Together these rows describe the final skill assignment for each task\. Section[4\.1](https://arxiv.org/html/2610.20926#S4.SS1)specifies aggregation; Appendix[B](https://arxiv.org/html/2610.20926#A2)gives the reference sources\.

## Appendix FCross\-Model Execution Analysis

We first examine how evolved skills help receivers when the developer’s final performance remains imperfect, then analyze unresolved failures\. The matched Luna comparison yields 11 gains, 23 ties, and seven losses among the 41 modified tasks\.

### F\.1Reusable workflow improvements and receiver execution

On six of Luna’s 11 improved tasks, Mini’s final mean reward is below 100\.0\. We examine Software Dependency Audit and Sales Pivot Analysis through 33 executions: 24 with final evolved skills, six with original skills on Luna, and three original\-skill Mini audit executions\. For each task, the final\-skill executions use the same frozen bundle across all four models\. Both tasks use binary execution rewards, so a mean of 33\.3 or 66\.7 denotes one or two successes out of three\.

A shared workflow error is corrected, while a downstream failure remains\.In the original\-skill evaluations, all three Mini executions and two of Luna’s three expand the scan to development dependencies and report 22 findings\. Luna’s remaining original execution keeps the intended scope but truncates a fix list\. With the final audit bundle,c01\(Appendix[I\.1](https://arxiv.org/html/2610.20926#A9.SS1)\), all 12 final executions across the four models keep the intended production scope and report three HIGH\-severity findings\. Mini rises from 0\.0 to 66\.7, and Luna from 0\.0 to 100\.0\. Mini’s failed final execution writes an incorrect CVSS score of 9\.8; the source\-priority rule requires 8\.1\. It also truncates a three\-version fix list to one version\. All three receivers preserve the required fields\. The remaining Mini error concerns a pre\-existing field rule, beyond the added row\-count and record\-key checks\.

The receiver must apply the correction to the actual input\.With the final Sales Pivot bundle,f07\(development in Appendix[I\.4](https://arxiv.org/html/2610.20926#A9.SS4)\), all Mini executions pass state validation, but two write summary tables without actual pivot objects\. Luna’s original\-skill executions create pivots but fail state validation by retainingOther Territories\. With the evolved bundle, Luna filters to the eight full state and territory names in the input, excludes four out\-of\-scope records, and builds pivots over the remaining 2,450 regions\. Its executions satisfy both requirements\. This requires adapting the skill’s whitelist example, which still uses abbreviations such asNSW\. GPT\-5\.4’s failed execution uses those abbreviations; its two successes retain full names\. GPT\-5\.5 builds pivots but retains the excluded category in all three executions\. These outcomes depend on adapting the guidance to the input’s state representation and completing both requirements\.

Reusable guidance can coordinate operations across saved artifacts\.The final Sales Pivot skill also specifies dependencies between operations: clean missing income markers before computing quartiles, and write, save, reopen, and check the source sheet before attaching pivot objects\. Mini’s successful final execution computes quartiles over valid income values and maps the labels back to the merged data\. Section[4\.4](https://arxiv.org/html/2610.20926#S4.SS4)describes a successful Luna execution that follows the same sequence of saving, reopening, and checking through its own builder\.

Table 12:Luna tasks with observed transfer losses\. Seven modified tasks lose reward under evolved bundles\. Values are mean rewards \(%\);Change \(pp\)uses unrounded measurements\.
### F\.2Unresolved failures in Luna reuse

For Luna’s seven loss tasks, we compare three original\-skill and three evolved\-skill executions per task \(42 total\), together with the corresponding edits\. Table[12](https://arxiv.org/html/2610.20926#A6.T12)reports mean rewards\.

Optimization and directory contracts\.Energy and Grid each fall from two successes to none after independently replacing demandPdP\_\{d\}withPd\+GsP\_\{d\}\+G\_\{s\}, motivated by passing development evidence\. All six evolved executions apply this rule and pass schema and feasibility checks, but fail cost optimality in the high\-cost direction\. The benchmark usesPdP\_\{d\}\. The four passing original executions usePdP\_\{d\}; the remaining original execution of each task fails the cost check in the opposite direction\. Passing development evidence therefore did not establish that the revised demand rule was correct\.

Organize also falls from two successes to none\. Every evolved execution places correctly classified documents in destinations below the source directory,/root/papers/all/\. Its inventory, manifest, and exact\-once audits check the destination tree, whereas the verifier treats documents remaining below the source directory as unsorted leftovers\. The failed original execution makes the same choice, while the two passing ones use sibling destinations under/root/papers/\. The stronger self\-check leaves source\-directory clearance unspecified\. For Mini, Energy improves, while Grid and Organize tie their original\-skill scores\.

Saving the solution\.Civ6’s zero\-reward executions \(one original and two evolved\) find competitive legal layouts but continue enumeration without saving the required JSON before timeout\. Two other executions time out with valid outputs and receive 0\.95 normalized adjacency reward\. Missing artifacts account for the loss; the submitted layouts remain competitive\. The evolved workflow lacks a requirement to save each new best layout before continuing\. Its helper checks serialization and supplied\-score sums but omits the stated adjacency recomputation\. The observed missing\-output failures occur before an artifact is available for that check\.

Computation and perception\.Drone’s three low\-scoring executions use quintic profiles; three high\-scoring executions use clamped cubic interpolation\. Although the task text permits broader continuous trajectories, the original skill and oracle specify cubic interpolation\. The mismatch between these quintic trajectories and the oracle’s cubic interpolation explains 13 common command failures, including all failures in the low\-scoring original execution\. The revised skills target hover and completion, but hover succeeds throughout, the cubic\-planning skill is unchanged, and the revision does not resolve the choice between these planners\. Scoring counts successful commands individually, so all enclosing pytest groups pass even at low reward\.

JPG workbooks cover all 22 receipts with correct structure and monetary values\. Five failed executions misread one or two month tokens on the same receipts across conditions; the errors do not involve swapping day and month\. Replacing multi\-pass guidance with deterministic parsing leaves perception errors despite image review; parsing cannot verify recognition accuracy\.

The only failed R2R execution passes the structure, linearization, safety, and local artifact checks, but fails the verifier’s settling\-time test\. The agent’s local settling\-time metric uses an all\-channel absolute\-error band and a post\-step time origin; the verifier uses the changed roller’s band relative to reference\-step magnitude and the logged starting time\. These metric conventions also differ in passing executions under both skill conditions; the task specifies a settling target without fully defining the measurement convention\. Applying the verifier’s rule to the failed execution’s saved log reproduces its settling\-time failure\.

Saturation and unresolved tasks\.Beyond these seven loss cases, the 23 tied tasks comprise 16 perfect, four zero, and three partial comparisons with two successes per condition\. The zero ties are Data\-to\-D3, Paratransit Routing, Manufacturing Codebook Normalization, and Quantum Numerical Simulation; only the first two improved on Mini\. The partial ties are React Performance Debugging, SEC Financial Report, and Shock Analysis Supply; only SEC improved on Mini\. Perfect ties leave no room for observed gains; other ties include failures under both skill conditions\.

### F\.3Additional receiver evidence

Shared tasks, different failures\.GPT\-5\.5 also loses reward on Grid, Organize, and Civ6\. Grid has two failures at a 0\.0111% cost gap against a 0\.01% limit, yet every execution readsPd\+GsP\_\{d\}\+G\_\{s\}, including the passing one\. That instruction alone does not determine reward\. Organize’s failed executions misclassify one paper; only one also violates directory clearance, and another execution passes\. Civ6 has one missing\-output timeout and two 0\.95 outcomes, showing the same checkpoint weakness with a different frequency\.

Saving an improved solution\.All three Paratransit GPT\-5\.5 executions time out, submitting reports serving 0, 440, and 442 trips against a minimum of 443\. Two find and locally check better feasible reports under other filenames but fail to save them as the graded output\. The remaining execution continues optimization after saving a below\-threshold report\. These executions expose the need to save the best available solution as the graded output\.

Answer and execution checks\.GRPO’s two 0\.75 GPT\-5\.5 outcomes fail only the epsilon source\-pattern check\. Their patches bind a named constant to10−410^\{\-4\}, whereas the verifier requires the denominator to containstd\_grouped\_rewardsplus an inline numeric literal\. The passing execution uses that literal\. The quarter\-point loss reflects the code pattern despite the correct numeric value\.

Lake Warming’s two GPT\-5\.5 failures attribute the dominant contribution to Human; the passing execution reports Heat\. The attribution calculation differs across these outcomes, while formatting does not explain the difference\. React’s two failures pass ten other checks but violate the external\-API check because checkout times of 8 and 11 ms fall below the 400 ms minimum\. SEC’s three normally completed executions fail answer checks on the newly acquired\-position list in one case and on the number of reported holdings in two\.

## Appendix GChanges to Skill Artifacts

Table[13](https://arxiv.org/html/2610.20926#A7.T13)summarizes changes between original and deployed skill bundles\. Counts cover the files supplied to the agent, distinguishing core instructions from executable support\. Each skill package is counted separately for each task that uses it\. Across all 86 tasks, 62 of 228 task\-skill pairs change \(27\.2%\); within the 41 modified tasks, these 62 changes occur among 125 pairs \(49\.6%\)\. Denominators use the union of original and evolved packages\. Per\-task means, medians, and correlations use the 41 modified tasks, whose total source length grows by 1\.5%\.

Table 13:Changes from original to deployed skill bundles\. Core instructions become shorter on average while support scripts grow\. The first two rows cover all 86 tasks; remaining statistics use the 41 modified tasks\. Means and medians are per task\. Packages and files are counted separately for each task; 228 counts the union of original and deployed packages\. Line edits = insertions \+ deletions; net change = insertions−\-deletions\.ρ\\rhois Spearman correlation\.aAll added support files are scripts \(17 Python, one shell\)\.

Line edits count insertions plus deletions; net length change counts insertions minus deletions\. Files are matched by skill and relative path within each task; a rename counts as deletion plus addition\. Core instructions, scripts, and references are classified by their bundle paths\.

## Appendix HSearch Histories and Development Paths

We trace each frozen non\-root deployment back to its recorded search root along primary\-parent links\. The 41 task histories contain 1,283 candidate entries: 971 retained candidates and 312 duplicate entries removed by deduplication\. Duplicate entries are not additional independently evaluated candidates\. GPT\-5\.4 Mini performs diagnosis and writing throughout, including the 126 edits on final adopted paths\. These search\-process summaries cover the 41 modified tasks and exclude the 45 tasks that retain their original bundles\.

Evaluation coverage\.Decision rankings contain 705 means from three executions each, covering 529 candidate versions and 176 incumbent comparisons\.

Actual expansion and later deployment\.An expansion event identifies a task, generation, and parent with at least one observed child\. It is non\-incumbent when that parent differs from the deployment incumbent entering the generation\. This label records parent identity within a run; the survivor\-selection rule determines whether development is non\-elitist\. We count an event as leading to deployment when one of its generated children, or a descendant of that child, is selected for deployment in that generation or a later one\. Eligibility alone does not count as an expansion event\.

There are 418 expansion events, of which 289 use non\-incumbent parents\. Of these 289 events, 76 \(26\.3%\) have a descendant later deployed, and 61 \(21\.1%\) enter the final deployed ancestry; each outcome spans 30 tasks\. Among the 76, 74 lead to a descendant accepted through a reward improvement or tie\-break\. The remaining two lead to Python\-to\-Scala deployments without a complete incumbent comparison\. Deployment decisions are available for 178 of 180 groups defined by task and generation; unavailable decisions are not assigned outcomes\.

Continuation from strictly lower\-scoring candidates\.We measure relative reward against the deployment selected in each candidate’s birth generation\. This comparison is distinct from a candidate’s reward change relative to its own parent\. Among the 529 fully measured candidate versions, 247 have a three\-execution mean more than 0\.05 percentage points below the deployment selected in the same generation’s decision ranking\. Both scores use that ranking’s complete three\-execution records, including the freshly measured incumbent when it remains deployed\. Of these lower\-scoring candidates, 163 are later expanded; 45 \(27\.6%\) have a subsequently deployed descendant across 20 tasks, and 35 lie on a final deployment’s primary\-parent path\.

For the 45 observed continuations, primary\-parent distance to the first deployed descendant is one, two, three, or four edits in 27, 12, three, and three cases\. Thus 18 candidates \(40\.0%\) are separated from that descendant by at least two edits\. The corresponding generation gaps are one, two, three, or four in 26, 12, four, and three cases\. Revisiting an earlier node can skip generations along a single edge\. These are candidate\-level counts; several candidates can belong to one task or ancestry\.

Figure[4](https://arxiv.org/html/2610.20926#S4.F4)a traces the final Gravitational Wave Detection path from root throughb05\(0\.0\),c05\(0\.0\),d01\(33\.3\),e04\(0\.0\), andf04\(100\.0\)\. Data\-to\-D3 follows the pathb05\(33\.3\),c02\(66\.7\),d03\(33\.3\),e01\(100\.0\), andf01\(100\.0\)\. Generation three retainsc04for deployment while keepingd03eligible, then generation four deploysd03’s childe01\. These are birth\-generation three\-execution means in percent\. Both tasks raise final Mini reward from 0\.0 to 100\.0\.

Table 14:Parent eligibility beyond the historical top three\. Both searches retain a parent despite three higher\-scoring saved versions\. Scores give each version’s latest complete three\-execution mean \(%\) at that generation\. Independent final Mini evaluation gives both tasks 100\.0, up from 0\.0\.Parent eligibility differs from a global score ranking\.Among the saved candidates recorded at each stated generation, both searches retain a parent outside the historical top three \(Table[14](https://arxiv.org/html/2610.20926#A8.T14)\)\. Descendants of both retained parents enter the final deployed ancestry\.

Explicit information sources outside the final path\.Search produces 75 graft candidates: 64 retained and 11 removed as duplicates\. Eight retained graft nodes lie on final primary\-parent paths across six tasks\. Their proposals identify donors outside their tasks’ final paths; six donors were not selected for deployment in their own birth generation\. Each proposal distinguishes the primary parent from the source of donated material\. Appendix[I](https://arxiv.org/html/2610.20926#A9)follows the donated procedure through the revised Sales Pivot bundle and its executions\.

Table 15:Adopted lineage depth\. The 41 adopted lineages have mean depth 3\.07 and median depth three\. Depth counts primary\-parent edits from the recorded root to deployment\. It can be smaller than the generation number when a search revisits an earlier node\.Adopted lineage depthTasks162133546511Table 16:Recorded edit operators\. Counts cover the 41 experiment trees that produced adopted non\-root skills\.Candidatescounts retained candidate versions,Edits on final pathscounts adopted ancestry edges, andFinal editscounts the 41 edges producing the final deployed bundles\. These frequencies do not measure operator success rates\. The 971 candidate records and 126 edits on final paths all record Mini as both diagnosis and writing model\.aThe name distill denotes a skill\-edit operator, not teacher\-model distillation\.

Depth and edit operators\.The 41 adopted lineages contain 126 edits, with 35 lineages having at least two edits and 11 reaching depth five\. Mean depth is126/41=3\.07126/41=3\.07\.

Tables[15](https://arxiv.org/html/2610.20926#A8.T15)and[16](https://arxiv.org/html/2610.20926#A8.T16)summarize complementary aspects of this history\. Depth counts successive edits from the recorded search root to the deployed bundle along primary\-parent links\. Operator counts cover the 971 retained candidates and the edits on adopted paths, excluding the 312 duplicate entries\.

The deployed Wyckoff and Exoplanet lineages begin with root children generated after the first generation\. The Paratransit and Suricata lineages contain parent links that skip the immediately preceding generation\. Primary\-parent depth can be smaller than generation number because the search can revisit a bundle from before the immediately preceding generation\.

Non\-incumbent parents appear on final revision paths\.Of the 87 adopted edits generated after generation one, 61 \(70\.1%\) use a parent other than the preceding generation’s deployment incumbent, spanning 30 of the 41 adopted lineages \(73\.2%\)\. Every chosen parent belongs to the eligible set\. This final\-path statistic has a different denominator from the 289 non\-incumbent expansion events above; it includes two later\-generation edits whose parent is root\.

Among the 29 tasks with improved final Mini reward, 22 \(75\.9%\) use a non\-incumbent parent on the final path, accounting for 46 of their 65 post\-generation\-one edges\. The corresponding counts for the 12 tied tasks are eight tasks and 15 of 22 edges\. Sixteen of the 29 improved\-task final paths pass through an expanded candidate that scored strictly below its birth\-generation deployment under the definition above\. We use the same 0\.05\-percentage\-point display band as in the endpoint analysis\.

Suricata retainsb05for deployment after the generation\-two hold, while reselectingb01and root as parents\. The reselectedb01produces the non\-incumbentd05in generation three, and the final ancestry continues throughe06andf01\. Thed05edit distills a validate\-then\-compose workflow for HTTP\-field rules\. Independent final rewards remain 0\.0% for Mini and 100\.0% for Luna\.

Paratransit provides a final\-path revisit with a positive final Mini difference \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)c\)\. Generation four retainsd06for deployment and reselectsd06,c05, andd04as parents\. Generation\-fivef08is a child of generation\-threed06and becomes the final selected bundle\. The complete path is root→\\tob07→\\toc03→\\tod06→\\tof08;d06also has generation\-four children, includinge03\. Mini’s independent final score rises from 0\.0 to 33\.3, while Luna remains at 0\.0\.

Later execution evidence guides revisions of earlier bundles\.Wyckoff returns to the original bundle after generation two, producingd08in generation three and finale03in generation four \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)d\)\. The root\-to\-d08edge bypasses two intervening generations, while the edit uses evidence from executions performed after the root bundle was created\. Its editing input explicitly quotes the generation\-oneb06execution’s batch analysis of mounted CIF files\. The new skill adds a file\-based recipe, Wyckoff\-letter grouping, and structured output requirements\. Ind08’s second execution, the agent reads the added guidance and runs a batch analysis over those files\. The revision makes the task workflow explicit using the symmetry\-analysis method already available in the original reference material\. Thed08search rewards are 100\.0, 82\.0, and 100\.0, giving a 94\.0 mean\. Generation three deploysd04but retainsd08as a parent, whose childe03is selected in generation four\. Independent evaluation of the complete finale03bundle raises Mini from 75\.7 to 100\.0, while Luna’s reward remains unchanged at 100\.0\.

Table 17:Skill\-revision cases and final bundle performance\.Scores show original→\\toevolved reward \(%\) for complete bundles\. These cases illustrate recorded revisions; the scores do not isolate individual edit effects\. Evaluation settings follow Section[4\.1](https://arxiv.org/html/2610.20926#S4.SS1)\.
## Appendix IConcrete Examples of Skill Evolution

These cases connect execution contrasts, proposals, and bundle changes\. Table[17](https://arxiv.org/html/2610.20926#A8.T17)reports representative final task scores\. The generation\-level scores below are means over three search\-time executions, expressed as percentages and distinct from final evaluations\. Mini performs execution, diagnosis, and editing throughout these development runs\.

### I\.1A persistent failure moves the repair upstream

Software Dependency Audit followsroot→b06→c01\\text\{root\}\\rightarrow\\texttt\{b06\}\\rightarrow\\texttt\{c01\}\. Root andb06each have a mean reward of 66\.7%\. The failed root execution expands the scan with\-\-include\-dev\-depsand writes 22 findings; the two passing executions use the default production\-dependency scope and write three rows\. Theb06reporting edit keys vulnerabilities by package, version, CVE identifier, and severity, removes duplicates, and checks the saved CSV against the scan\-derived set\. Its failed execution deduplicates the 22 rows and verifies their count and key uniqueness, but still expands the scan\. Both failures pass the structural tests but fail the content comparison\. Checking the report against an expanded scan leaves the required scan scope unverified\.

Thec01proposal therefore moves the repair upstream\. Its scanning instructions prohibit\-\-include\-dev\-depswhile leavingb06’s reporting skill unchanged\. All three executions use the production scope, write three rows, and check the saved count and key set, yielding 100\.0% mean reward\. The revision makes explicit the flag’s omission from the original command example\.

### I\.2Workflow consolidation on a retained branch

In Gravitational Wave Detection, generation four deployse01but retainse04as a parent \(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)a; full path in Appendix[H](https://arxiv.org/html/2610.20926#A8)\)\. Althoughe04’s mean reward is 0\.0%, each execution passes eight of nine tests and fails only the expected numerical\-result comparison\. In all three runs, waveform generation and matched filtering use the 15 Hz conditioning cutoff, violating the 20 Hz cutoff specified in the separate search skill\.

Thef04proposal consolidates the two\-skill bundle\. It moves the search, final write, and readback procedure into theconditioningdocument, then removes the separatematched\-filteringdocument\. The helper file is unchanged\. The merged workflow specifies 15 Hz conditioning alongside the 20 Hz cutoff for waveform generation and matched filtering\.

With the consolidated bundle, all threef04runs use 20 Hz for search while retaining 15 Hz conditioning\. All three executions of each bundle,e04andf04, write and reread a correctly structured CSV\. Thef04runs pass every test and win deployment with a 100\.0% mean reward, compared with 66\.7% fore01’s fresh batch\. Workflow consolidation accompanies a change in the executed computation, and a zero\-reward parent produces the next deployed version\.

### I\.3Helper repair and direct workflow execution

In PDDL Airport Planning, root’s search\-time mean reward is 0\.0%\. The generation\-oneb06proposal responds to failed plan syntax by requiring writer\-based serialization and parsing the saved plan against the same problem inputs\. The initial edit omits its promised helper; validation feedback prompts a repair that adds the executable and its invocation inSKILL\.md\. Although the repaired bundle now includes the helper, the agent cannot locate it in the firstb06execution\. The agent implements the prescribed writing, parsing, and validation workflow directly through the planning library\. The saved plan passes official verification\.

### I\.4A non\-deployed sibling contributes an executed procedure

Sales Pivot Analysis separates the value of a bundle for deployment from the value of its individual procedures\. Generation one deploysb04, which revises spreadsheet construction and has a mean reward of 100\.0%\. Siblingb05revises PDF extraction and has a mean reward of 33\.3%\. Its first two executions fail pivot\-configuration checks while passing state validity\. Thec04proposal selectsb04as parent andb05as donor, citing the donor’s successful third execution\.

Thec04bundle containsb05’s continuation\-aware PDF procedure, including header handling and state filtering, while leavingb04’s spreadsheet skill unchanged\. Its two passing executions use header\-aware extraction and retain full state names\. The third converts those names to abbreviations and fails state validity, yielding a search\-time mean reward of 66\.7%\.

Generation two retainsb04after the historical\-score check but selectsc04for further development\. The final bundle descends fromc04; its primary\-parent path therefore includes the revision that imports the procedure fromb05\(Figure[4](https://arxiv.org/html/2610.20926#S4.F4)b\)\.

Procedural compression retains perfect scores\.Erlang SSH’s generation\-oneb09prunes a 485\-line penetration\-testing guide to 31 lines focused on reproducing an unauthenticated channel\-open attempt and verifying rejection, adds no script, and preserves perfect final scores on Mini and Luna\.

## Appendix JProposer and Skill Editor Prompt Excerpts

These excerpts from recorded Mini calls for Software Dependency Audit show how the Proposer compares executions and specifies an edit for the Skill Editor\. For brevity, some background and runtime details are omitted; operational rules and output formats are retained\. Complete templates accompany the source files\. Here,distillencodes procedures from executions in the bundle\.

Listing 1:Generation\-one Proposer excerpt\. The prompt provides root execution evidence, instructions for diagnosing failures, and the proposal output format\.Differentialevolutionofaskilllibrary,generation1fromroot\(1\-taskpool\)\.Evidence:freshrootrolloutsbelow\(failuresandbehavior\)\.

\[ROOT\_EXECUTION\_EVIDENCE\]

\#\#Alreadyproposedthisgeneration\(doNOTrepeatornear\-duplicate\)

\[PREVIOUS\_PROPOSALS\]

\#\#ProposalsREJECTEDsofarandwhy\(eachcostaroundandproducednochild\)

\[REJECTION\_FEEDBACK\]

\#\#Allowedoperators\(oneperchild\)

prune/rewrite/extend/create/consolidate\.EveryoneofthesewritesaCOMPLETESKILL\.md,soanyofthemcanchangethedescription,thebody,orboth,and,wherethescriptssectionbelowappears,thescript\-capableonescanshipafileunderscripts/inthesamechild\.Choosebywhattheevidencesaysisbroken;thereisnodefaultoperator\.createONLYifnohostskillexists:narrowdescriptionnamingaconcretetaskdomain,max2pergeneration\.

consolidateMERGES:thetargetskillsurvivesunderitsownnameandeveryOTHERskillinthelibraryisfoldedintoitanditsdirectorydeleted\.Themergeddescriptionmuststillfirewherevereachfoldedskillfired\-\-itisrejectedifitdropsthetriggertermsnosurvivingskillcarries\-\-andthemergedbodymustteacheachfoldedprocedureinoneorderanagentcanfollow,notconcatenatethem\.Onlythesecansurviveafoldhere:cvss\-score\-extraction,trivy\-offline\-vulnerability\-scanning,vulnerability\-csv\-reporting\.

AchildmayalsoshipEXECUTABLEfiles:rewrite,extend,create,consolidatemayaddnewfilesunderthetargetskill’sscripts/,andeveryoneofthemisruninsidethistask’sowncontainerbeforethechildisaccepted\.AimadirectiontherewheretheevidenceshowstheagentdoingsomethingEVERYRUNthataprogramcoulddoonceandbetter:derivingavalueitcannotlookup,orconfirmingaresultitcannototherwiseobservebeforesubmitting\.Prosecannotsupplythefirstanddoesnothelpwiththesecond\.Ahelperthatonlyrestatesthetextcanmerelybreak\.ShippingonemeanswritingthefileANDthesentencethatmakesitreachable;thoseareoneedit,nottwo\.

\#\#Existingskills

\[EXISTING\_SKILL\_NAMES\]

\#\#Consistencyreview\(advisory,notacontentlock\)

Anunmetgradernounisnotaneditdirection\.Ifagradersaysanobjectisinvalid,

missing,extra,orunreachable,firstusethevisiblerunstatetodecidewhetherthe

losingrunaddedit,omittedit,ortransformeditwrongly\.Donotturnthenounalone

intoaninstructiontoaddorpreservethatobject\.

Agent\-visibleexecutablesourceandastate/actionfromarunthatPASSEDareconsistencyevidencewhichthediagnosismustreconcile\.Theyarenotimmutablecontent:aproposalmayreplaceexistingbehaviorwhenthefailing/passingcontrastexplainswhyandstatesthenarrowerinvariantthatshouldreplaceit\.Foreachproposal,quotetherelevantexecutablelinewhenthetargetskillhasoneintheevidenceblock,andquoteaconcretesuccessfulstate/actionwhenapassingrunisshown\.Arecoveredexception,acommandlinebyitself,or"Processexitedwithcode0"isnotsuccessful\-stateevidence\.Thenstatespecificallywhetherthedirectionpreserves,narrows,ordeliberatelyreplacesthequotedbehaviorandwhythatisconsistentwiththemeasuredcontrast\.Thesefieldsareretainedforaudit\.Missingorimperfectquotationisadvisoryandneverdiscardsaproposal;officialevaluationremainstheonlyarbiteroftheedit\.UseNONEonlywhenthecorrespondingkindofevidenceisgenuinelyabsent\.

\#\#Howtoreadtheevidence,decide,andrank

Howtoreadtheevidence\.Thegrader’sverdictblocksaboveareWHATdecidesthescore:arunscoreszerobecausearequirementwasstillstandingwhenitended,andeachrequirement’sshown/lost/survivedcountsarebesideit\.TherolloutblocksareWHY\-\-thatiswhereyoudotheactualthinking\.Theyarelongandfullofthingsgoingwrong,andmostofitistheagentfindingitsownerrorsandfixingthem;ablocktaggedascomingfromareplicatethatPASSEDisproofitserrorswererecoverable\.Anerrorarunrecoveredfrom,orarequirementthatalsoappearsinrunsthatpassed,isnotwhatlosesthetask,howeveroftenitappears\.

AcausehastoEXPLAINtheverdict,notappearinit\.Acauseyoucannottracetotheverdictatallisadetour;onetheverdictnevernamesbutthatexplainsitisusuallytherealone,andyoureachitonlythroughtheruns\.

TheverdictisforYOU,notfortheagent\.Theagentneverseestheverifier;itgradesaftertheagenthasfinished\.SoadirectionmustbeactionableonwhattheagentCANobservebeforeitfinishes:thesourceitwasgiven,thefilesinitsworkingtree,theoutputofacommanditcanrun\.AndtheverdictLOCATEStheproblem;itisnotananswertohandover\.Aneditthattranscribesthenamesthegradermentionshascachedthisinstance’sanswer\-\-theagentcanalreadyreachthem\.AimthedirectionathowtoOBTAINtherequiredsurface\.

WhatyouaredecidingiswhatwouldmakethistasksucceedRELIABLYratherthansometimes\.Youhaveroot’sreplicatesabove;thewriterdownstreamseesonedirectionandthefileitedits,sothejudgementhastobemadehereandstatedinthedirection,specificallyenoughthatsomeonewhocannotseetheserunscouldcarryitoutexactly\.SotheotherhalfofeveryfindingisthepaththatWORKED:atthepointafailingrunwentwrong,whatdidapassingrundoinstead?Wherethereplicatescontainapass/failpair,thatisthecleanestsource\-\-thelibrarytextwasidentical,sowhateverdiffersisthepath\.Quoteitinwhat\_passing\_runs\_did;"norunhasevergotpastthis"isarealanswer,anditlabelstheeditanuntestedguess,whichisworthknowing\.

Rankthecausesbeforeyoupropose\.StateeachcauseasaSENTENCEnamingamechanism\.RankbyhowmanyreplicatesLOSTbecauseofit,nothowmanyshowit,andreportbothplushowmanyshoweditandpassedanyway\.Ifacauseisoutsidewhatanyskilleditcouldaddress,stillrankithonestlyandmarkit"skill\_can\_address":false\-\-theattackquotabelowthenappliestothehighestcauseaskillCANaddress\.

Onesignaturedeservesprioritywheneveritappears:theGRADEDVALUEclusters\.WhenthenumberthegradercheckscomesoutinafewtightgroupsacrossrunsoftheSAMElibrary\-\-withinagroupthevaluesagreetofarmoreprecisionthanchancewouldallow\-\-thatisnotnoise:therunstookdifferentdeterministiccomputationpathsatsomestep,andwhichpatharuntakesdecidesthescore\.Diffarunfromeachclusteratthestepthatproducesthegradednumber,namethedivergenceasacause,andproposepinningthepassingpathasanexplicitprocedureintheskill\-\-theexactordering/normalization/formula,notthetargetvalue\.Rankitthroughthesameevidencerulesaseveryothercause\.

Atleast2ofyour6proposalsmustattackthefirst\-rankedaddressablecause\(causestiedatthetopscorecountasjointlyfirst\)andatleast3mustattackthetoptwo;therestarefree,andthatiswhereasmalloroddobservationbelongs\.

Beforecommittingtoeachproposal,checkitagainsttheruns:thecauseaccountsfortherunsthatlost;itisabsentfrom,orvisiblyovercomein,therunsthatpassed;andthedirectionisexecutablebysomeonewhocannotseetheseruns\.Aproposalthatfailsanyofthesewastesitschild\.

\#\#OutputstrictJSONonly

\{"loses\_the\_task":\[\{"cause":"<onesentencenamingtheMECHANISMthatlosesthetask;itmustEXPLAINthegrader’scomplaintandmaynamethingstheverdictnevermentions\>","skill\_can\_address":<falseONLYifnoskilleditcouldactonthiscause;omitotherwise\>,"runs\_showing":<howmanyrunsshowitatall\>,"runs\_affected":<ofthose,howmanyLOSTbecauseofit\-\-arunthatshoweditandstillpasseddoesnotcount\>,"also\_in\_passing\_runs":<howmanyshoweditandpassedanyway;0isastrongcause,alargenumbermeansyouarelookingatasymptom\>,"what\_passing\_runs\_did":"<thesteparunthatPASSEDtookatthissamepoint,quotedfromitstrace\-\-thecommand,thelineitwrote,theorderitdidthingsin\."norunhasevergotpastthis"isarealanswerandsaysthefixisnotyetintheevidence\>","evidence\_quote":"<verbatim\>"\},\.\.\.\],

"diagnoses":\[\{"parent":"root","operator":"prune\|rewrite\|extend\|create\|consolidate",

"target\_skill":"<name\>","direction":"<one\-lineedit\>",

"problem":"<whattheevidenceshows\>",

"failing\_test":"<thetestthatfailsinmostreplicates,byname\>",

"diverged\_at":"<thestepintheFAILINGrunthatproducedwhatthattestrejects\-\-thecommanditranorthefileitwrote,quotedfromthetrace\.Ifnoruneveropenedthisskillthereisnosuchstep:sayso,andthatabsenceisitselfthefinding\>",

"would\_change":"<whattheagentdoesdifferentlyatthatstep,andwhetheritcouldhaveknowntherethatitwaswrong\-\-ifitcouldnot,saywhatmakesitobservable\>",

"executable\_consistency\_quote":"<verbatimtaggedlinefromthistargetskill’sAGENT\-VISIBLEEXECUTABLEblock,orNONEonlyifabsent\>",

"passing\_consistency\_quote":"<verbatimtaggedsuccessfulstate/actionfromapassingrun,orNONEonlyifabsent\>",

"consistency\_check":"<whetherthisdirectionpreserves,narrows,orreplacesthequotedbehavior,andwhythemeasuredpass/failcontrastsupportsthatrelationship\>",

"ships\_file":"<trueifthiseditdeliversanexecutablefile\.Saytruewheneverthecauseisthatrunssubmitsomethingtheyneverchecked\-\-aninstructioncanonlysaywhattodo,acheckreportswhetheritwasdone\.Thecontractisthatadeclarationmeansafile:saytrueandthewritermustemititina===SCRIPTblock,andprosedescribingafileisnotthefile\>",

"attacks":"<theexact‘cause‘stringfromloses\_the\_taskthiseditisaimedat\>",

"evidence\_quote":"<verbatimfromevidenceabove\>"\},\.\.\.\]\}

Theyareaskedforbecauseaneditthatcannotsaywhereinarunitactshaschangedatopicandnotarun\.Answerthemfromthetrace,andanswerthemhonestly\-\-"norunopenedthisskill,sothereisnosuchstep"isarealanswerandpointsatthedescription;aninventedstepisworthlessthannone\.

Exactly6diagnoses,diversetargetsandoperators\.\[\.\.\.\]

Listing 2:Later\-generation Proposer excerpt\. The prompt supplies candidate comparisons, parent and donor constraints, and the proposal output format\.Differentialevolutionofaskilllibrary,generation2\.Breedingparents:b06,b01,b04\(lineagesofroot\+accumulatedsingleedits\)\.Evidencebelow:graderverdicts,thepreviousgeneration’spopulationcontrast,andrawruntraces\.

\[POPULATION\_EXECUTION\_EVIDENCE\]

\#\#Alreadyproposedthisgeneration\(doNOTrepeattheseornear\-duplicates\)

\[PREVIOUS\_PROPOSALS\]

\#\#EditsalreadyMEASUREDasnetlosses\(closelysimilardirections\-\-sameoperatorandtarget\-\-areauto\-rejected\)

\[PREVIOUSLY\_MEASURED\_LOSING\_EDITS\]

\#\#ProposalsREJECTEDsofarandwhy\(eachcostaroundandproducednochild\)

\[REJECTION\_FEEDBACK\]

\#\#Allowedoperators\(oneperchild\)

prune/rewrite/graft/extend/create/drop/narrow\_desc/consolidate\.Everyoneoftheseexceptnarrow\_descanddropwritesaCOMPLETESKILL\.md,soanyofthemcanchangethedescription,thebody,orboth,andrewrite/graft/extend/create/distill/consolidatecanalsoshipafileunderscripts/inthesamechild\-\-writingthatfileandthesentencethatmakesitreachableisONEedit,nottwo\.Choosebywhattheevidencesaysisbroken;thereisnodefaultandnopreferredoperator\.graftmaysourceasectionfromANYofb06,b01,b04intoabreedingparent\(namesource\_node\)\.createONLYifnohostexists:narrowdescriptionnamingaconcretetaskdomain,max2pergeneration\.

distill:encodethemethodawinningrunactuallyused\.YOUaretheonereadingthoseruns,sostatethemethoditselfin‘direction‘and‘would\_change‘\-\-thestepsinorder,thecommandsorchecksthatmattered,whatthelosingrunsdidinstead\.Theeditorplaceswhatyouwrite;itdoesnotseetherunsandcannotre\-derivethem\.

DELIVERYMECHANISM:beforethetaskinstructiontheagentisshownEVERYmountedskillasname\+fulldescription,permanently,andistoldthatataskclearlymatchingaskill’sdescriptionMUSTusethatskill\.Nothingishiddenbehindadirectorylisting\.Thedescriptiondecideswhetherthebodyiseveropened;thebodydecideswhattheagentDOES\-\-andthebodydeliversintwoforms,notone\.ItcanINSTRUCT,whichtheagentthenfollowsaswellasithappensto,oritcanSHIPAFILEUNDERscripts/andnamethecommandthatrunsit,whichreports\.Aninstructioncanonlysaywhattodo;achecksayswhetheritwasdone\.Whereacauseisthatrunssubmitsomethingunverified,thefileisthestrongereditandtheoperatorsrewrite/extend/graft/distill/consolidate/createcanallcarryone\.Thislibraryholds3skills:nothingisburiedandthereisnosiblingtowinattentionfrom,sotheonlyusefuldescriptioneditmakestheskillfiremorereliably\-\-ENUMERATEtheconcretetriggersofthetaskitserves,inthattask’sownvocabulary\.Takingtriggersurfaceaway\(narrowing,disclaimingwhattheskillisnotfor\)hasnocompetitortobeathereandcanonlystoptheskillfiring\.Theleverageisinthebody’sprocedure\.

dropREMOVESonewholeskillfromthelibrary\-\-proposeitwheretheevidenceshowsaskillactivelymisleadstheagentonthistask;theremovalisthechild’sentireedit\.narrow\_descrewritesONLYaskill’sdescriptionline,leavingthebodybyte\-identical;whatthatlineshouldthensayissetbytheDELIVERYMECHANISMparagraphabove,whichdependsonhowmanyskillsthislibraryholds\.

consolidateMERGESTHELIBRARYDOWN:thetargetskillsurvivesunderitsownnameandeveryOTHERskillofthatparentisfoldedintoitanditsdirectorydeleted\.Bothhalvesarecheckedmechanically,soproposeitonlywheretheyhold:themergeddescriptionmuststillfirewherevereachfoldedskillfired\(itisrejectedifitdropsthetriggertermsnosurvivingskillcarries\),andthemergedbodymustteacheachfoldedprocedureinoneorderanagentcanfollowratherthanconcatenatingthem\.Whichskillmaysurviveafold,byparent\-\-b06,b01,b04:cvss\-score\-extraction,trivy\-offline\-vulnerability\-scanning,vulnerability\-csv\-reporting\.

AchildmayalsoshipEXECUTABLEfiles:rewrite,graft,extend,create,distill,consolidatemayaddnewfilesunderthetargetskill’sscripts/,andeveryoneofthemisruninsidethistask’sowncontainerbeforethechildisaccepted\.AimadirectiontherewheretheevidenceshowstheagentdoingsomethingEVERYRUNthataprogramcoulddoonceandbetter:derivingavalueitcannotlookup,orconfirmingaresultitcannototherwiseobservebeforesubmitting\.Prosecannotsupplythefirstanddoesnothelpwiththesecond\.Ahelperthatonlyrestatestheskill’sproseaddsriskwithoutaddinginformation\.ShippingonemeanswritingthefileANDthesentencethatmakesitreachable;thoseareoneedit,nottwo\.

\#\#Existingskills

\[EXISTING\_SKILL\_NAMES\]

\#\#Consistencyreview\(advisory,notacontentlock\)

Anunmetgradernounisnotaneditdirection\.Ifagradersaysanobjectisinvalid,

missing,extra,orunreachable,firstusethevisiblerunstatetodecidewhetherthe

losingrunaddedit,omittedit,ortransformeditwrongly\.Donotturnthenounalone

intoaninstructiontoaddorpreservethatobject\.

Agent\-visibleexecutablesourceandastate/actionfromarunthatPASSEDareconsistencyevidencewhichthediagnosismustreconcile\.Theyarenotimmutablecontent:aproposalmayreplaceexistingbehaviorwhenthefailing/passingcontrastexplainswhyandstatesthenarrowerinvariantthatshouldreplaceit\.Foreachproposal,quotetherelevantexecutablelinewhenthetargetskillhasoneintheevidenceblock,andquoteaconcretesuccessfulstate/actionwhenapassingrunisshown\.Arecoveredexception,acommandlinebyitself,or"Processexitedwithcode0"isnotsuccessful\-stateevidence\.Thenstatespecificallywhetherthedirectionpreserves,narrows,ordeliberatelyreplacesthequotedbehaviorandwhythatisconsistentwiththemeasuredcontrast\.Thesefieldsareretainedforaudit\.Missingorimperfectquotationisadvisoryandneverdiscardsaproposal;officialevaluationremainstheonlyarbiteroftheedit\.UseNONEonlywhenthecorrespondingkindofevidenceisgenuinelyabsent\.

\#\#RankthecausesBEFOREyouproposeanything

\[PREVIOUS\_GENERATION\_LEADING\_CAUSE\]

ItsevidenceisSTILLPRESENTintheblockabove\.Rankitagain,orstateasitsownentrywhyitisnolongerwhatlosesthistask\-\-aneditlandedandworked,ortherunsmovedon\.Droppingitsilentlyistheonethingthatisnotananswer\.

Howtoreadtheevidence\.Thegrader’sverdict\(thefirstblockabove\)isWHATdecidesthescore:arunscoreszerobecausearequirementwasstillstandingwhenitended,andeachrequirement’sshown/lost/survivedcountsarebesideit\.ThetrajectoriesareWHY\-\-thatiswhereyoudotheactualthinking\.Theyarelongandfullofthingsgoingwrong,andmostofitistheagentfindingitsownerrorsandfixingthem;anerrorarunrecoveredfrom,orarequirementthatalsoappearsinrunsthatPASSED,isnotwhatlosesthetask,howeveroftenitappears\.

AcausehastoEXPLAINtheverdict,notappearinit\.Acauseyoucannottracetotheverdictatallisadetour;onetheverdictnevernamesbutthatexplainsitisusuallytherealone,andyoureachitonlythroughtheruns\.

TheverdictisforYOU,notfortheagent\.Theagentneverseestheverifier;itgradesaftertheagenthasfinished\.SoadirectionmustbeactionableonwhattheagentCANobservebeforeitfinishes:thesourceitwasgiven,thefilesinitsworkingtree,theoutputofacommanditcanrun\.AndtheverdictLOCATEStheproblem;itisnotananswertohandover\.Aneditthattranscribesthenamesthegradermentionshascachedthisinstance’sanswer\-\-theagentcanalreadyreachthem\.AimthedirectionathowtoOBTAINtherequiredsurface\.

WhatyouaredecidingiswhatwouldmakethistasksucceedRELIABLYratherthansometimes\.Youhavethereplicatesandcontrastsabove;thewriterdownstreamseesonedirectionandthefileitedits,sothejudgementhastobemadehereandstatedinthedirection,specificallyenoughthatsomeonewhocannotseetheserunscouldcarryitoutexactly\.SotheotherhalfofeveryfindingisthepaththatWORKED:atthepointafailingrunwentwrong,whatdidapassingrundoinstead?Wheretheevidencecontainsasame\-librarypass/failpair,thatisthecleanestsource\-\-thetextwasidentical,sowhateverdiffersisthepath\.Quoteitinwhat\_passing\_runs\_did;"norunhasevergotpastthis"isarealanswer,anditlabelstheeditanuntestedguess,whichisworthknowing\.

Rankthecausesbeforeyoupropose\.StateeachcauseasaSENTENCEnamingamechanism\.RankbyhowmanyrunsLOSTbecauseofit,nothowmanyshowit,andreportbothplushowmanyshoweditandpassedanyway\.Ifacauseisoutsidewhatanyskilleditcouldaddress,stillrankithonestlyandmarkit"skill\_can\_address":false\-\-theattackquotabelowthenappliestothehighestcauseaskillCANaddress\.

Onesignaturedeservesprioritywheneveritappears:theGRADEDVALUEclusters\.WhenthenumberthegradercheckscomesoutinafewtightgroupsacrossrunsoftheSAMElibrary\-\-withinagroupthevaluesagreetofarmoreprecisionthanchancewouldallow\-\-thatisnotnoise:therunstookdifferentdeterministiccomputationpathsatsomestep,andwhichpatharuntakesdecidesthescore\.Diffarunfromeachclusteratthestepthatproducesthegradednumber,namethedivergenceasacause,andproposepinningthepassingpathasanexplicitprocedureintheskill\-\-theexactordering/normalization/formula,notthetargetvalue\.Rankitthroughthesameevidencerulesaseveryothercause\.

BeingDRAWNtoavividfailureisfineandoftenuseful;beingdrawnentirelytoitisnot\.Atleast2ofyour6proposalsmustattackthefirst\-rankedaddressablecause\(causestiedatthetopscorecountasjointlyfirst\)andatleast3mustattackthetoptwo;therestarefree,andthatiswhereasmalloroddobservationbelongs\.

Beforecommittingtoeachproposal,checkitagainsttheruns:thecauseaccountsfortherunsthatlost;itisabsentfrom,orvisiblyovercomein,therunsthatpassed;andthedirectionisexecutablebysomeonewhocannotseetheseruns\.Aproposalthatfailsanyofthesewastesitschild\.

\#\#OutputstrictJSONonly

\{"loses\_the\_task":\[\{"cause":"<onesentencenamingtheMECHANISMthatlosesthetask;itmustEXPLAINthegrader’scomplaintandmaynamethingstheverdictnevermentions\>","skill\_can\_address":<falseONLYifnoskilleditcouldactonthiscause;omitotherwise\>,"runs\_showing":<howmanyrunsshowitatall\>,"runs\_affected":<ofthose,howmanyLOSTbecauseofit\-\-arunthatshoweditandstillpasseddoesnotcount\>,"also\_in\_passing\_runs":<howmanyrunsshoweditandpassedanyway;0isastrongcause,alargenumbermeansyouarelookingatasymptom\>,"what\_passing\_runs\_did":"<thesteparunthatPASSEDtookatthissamepoint,quotedfromitstrace\-\-thecommand,thelineitwrote,theorderitdidthingsin\."norunhasevergotpastthis"isarealanswerandsaysthefixisnotyetintheevidence\>","evidence\_quote":"<verbatim\>"\},\.\.\.\],

"diagnoses":\[\{"parent":"b06\|b01\|b04","operator":"prune\|rewrite\|graft\|extend\|create\|drop\|narrow\_desc\|distill\|consolidate",

"source\_skill":"<forgraftONLY,optional:donorSKILLnamewhentransplantingasectionintoadifferentskill\>",

"target\_skill":"<name\>","source\_node":"<forgraftONLY:theLIBRARYidtocopythesectionFROM,oneofthecomparisonsetabove\-NOTaskillname;omitotherwise\>",

"task":"<OPTIONAL:thepooltaskthiseditaimstoimprove,exactlyasnamedintheevidence;whengiven,theeditorreceivesthattask’sinstructiontext\>",

"direction":"<one\-lineedit\>","problem":"<whattheevidenceshows\>",

"failing\_test":"<thetestthatfailsinmostreplicates,byname\>",

"diverged\_at":"<thestepintheFAILINGrunthatproducedwhatthattestrejects\-\-thecommanditranorthefileitwrote,quotedfromthetrace\.Ifnoruneveropenedthisskillthereisnosuchstep:sayso,andthatabsenceisitselfthefinding\>",

"would\_change":"<whattheagentdoesdifferentlyatthatstep,andwhetheritcouldhaveknowntherethatitwaswrong\-\-ifitcouldnot,saywhatmakesitobservable\>",

"executable\_consistency\_quote":"<verbatimtaggedlinefromthistargetskill’sAGENT\-VISIBLEEXECUTABLEblock,orNONEonlyifabsent\>",

"passing\_consistency\_quote":"<verbatimtaggedsuccessfulstate/actionfromapassingrun,orNONEonlyifabsent\>",

"consistency\_check":"<whetherthisdirectionpreserves,narrows,orreplacesthequotedbehavior,andwhythemeasuredpass/failcontrastsupportsthatrelationship\>",

"ships\_file":"<trueifthiseditdeliversanexecutablefile\.Saytruewheneverthecauseisthatrunssubmitsomethingtheyneverchecked\-\-aninstructioncanonlysaywhattodo,acheckreportswhetheritwasdone\.Thecontractisthatadeclarationmeansafile:saytrueandthewritermustemititina===SCRIPTblock,andprosedescribingafileisnotthefile\>",

"attacks":"<theexact‘cause‘stringfromloses\_the\_taskthatthiseditisaimedat\>",

"evidence\_quote":"<verbatimfromevidenceabove\>"\},\.\.\.\]\}

Theyareaskedforbecauseaneditthatcannotsaywhereinarunitactshaschangedatopicandnotarun\.Answerthemfromthetrace,andanswerthemhonestly\-\-"norunopenedthisskill,sothereisnosuchstep"isarealanswerandpointsatthedescription;aninventedstepisworthlessthannone\.

Exactly6diagnoses,diversetargetsandoperators,eachgroundedintheevidenceabove\.\[\.\.\.\]

### J\.1Skill editing from a proposal

The Skill Editor receives the proposal, parent skill, and task context\. The excerpt below illustrates a rewrite call; the current skill is represented by an input slot\.

ApplyONErewriteedittoanagentskill\(SKILL\.mdwithYAMLfrontmattername/description\+markdownbody\)usedbyafrozenCLIcodingagentonthesetasks\-\-software\-dependency\-audit\.

\#\#Theedittomake\-\-aspecificationtoexecute,nottexttocopy

\{

"operator":"rewrite",

"target\_skill":"trivy\-offline\-vulnerability\-scanning",

"direction":"Rewritethescaninstructionstokeepthisauditon‘/root/package\-lock\.json‘withthedefaultofflineproduction\-dependencyscopeandexplicitlyreject‘\-\-include\-dev\-deps‘\.",

"problem":"Somerunsbroadenthevulnerabilityuniverseandwriteadev\-inclusiveCSVthatdoesnotmatchthegrader’sexpectedset\.",

"failing\_test":"test\_outputs\.py::TestSecurityAuditTask::test\_csv\_matches\_ground\_truth",

"diverged\_at":"trivyfs/root/package\-lock\.json\-\-formatjson\-\-output/root/trivy\_report\_dev\.json\-\-scannersvuln\-\-skip\-db\-update\-\-offline\-scan\-\-include\-dev\-deps\-\-cache\-dir/root/\.cache/trivy",

"what\_should\_happen\_there":"Atthescancommand,theflagchoiceisdirectlyobservable,sotheagentshouldomitthedev\-depsflagandcontinueonlywiththeproduction\-onlyreport\.",

"decided\_from\_this\_line\_in\_the\_runs":"trivyfs/root/package\-lock\.json\-\-formatjson\-\-output/root/trivy\_report\_dev\.json\-\-scannersvuln\-\-skip\-db\-update\-\-offline\-scan\-\-include\-dev\-deps\-\-cache\-dir/root/\.cache/trivy",

"executable\_consistency\_evidence":"\[\.\.\.\]",

"successful\_state\_consistency\_evidence":"trivyfs/root/package\-lock\.json\-\-scannersvuln\-\-offline\-scan\-\-skip\-db\-update\-\-cache\-dir/root/\.cache/trivy\-\-formatjson\-o/root/trivy\_report\.json",

"diagnosis\_consistency\_check":"NarrowsthecommandpathtotheobservedpassingscanwhilepreservingtheofflineTrivyworkflow\."

\}

Objective:thiseditisjudgedononething\-\-whetherthefrozenagentpassesthistask’sowntestsmoreoften,repeatablyacrossindependentrunsratherthanoncebyluck\.WordingthatdoesnotchangewhattheagentDOESscoresnothing\.

Yourreaderisanagentdoingthistask\.Thediagnosisupstreamreadtherunsanddecidedwhatthatagentneeds;thedecisionisthedirectionabove,andyourjobistomakeitrealinthefile\.Theformisyours\-\-aprocedure,arule,acheckitruns,afactitwasmissing\-\-exceptwherethedirectionitselfnamesone\.

Whenthespecificationcontainsexecutable\_consistency\_evidenceorsuccessful\_state\_consistency\_evidence,reconcilethecompletededitwiththem\.Theyareevidence,notfrozencontent:preservethemwhentheydescribetheworkinginvariant,ordeliberatelyrevisethemonlywherethesupplieddiagnosis\_consistency\_checkexplainsthefailing/passingboundary\.

Askillreachestheagentintwolayers\.TheDESCRIPTIONisalwaysintheagent’scontextanddecideswhetherthebodyiseveropened\.TheBODYthendeliversoneoftwothings:instructions,whichtheagentfollowsaswellasithappensto,orafileitruns,whichcanreportwhethertheworkwasactuallydone\.Thedirectionandproblemabovetellyouwhichlayerthiseditactson\.

Threerulesforwhatgoesinthebody:

\-TheagentNEVERseestheverifier;itgradesaftertheagenthasfinished\.Neverwriteanythingtriggeredbygraderoutput\(’iftheverifierreportsX\.\.\.’\)\.Writetriggerstheagentcanobservebeforefinishing:thesourceitwasgiven,thefilesinitsworkingdirectory,theoutputofacommanditcanrunitself\.Statingarequirementasafactaboutthetaskisfine\.

\-TeachhowtoOBTAINwhatthetaskrequires,notwhatitis\.’Readthespecthetaskmountsandmatchthenamesitimportsbeforewritingtheimplementation’transfers;atableenumeratingthistask’sownnamesisacachedanswerthattransfersnowhere\.Oneortwonamesasanexampleoftheshapeisfine\.

\-Everylinemustactonthecausethedirectionnames\.Generalprocedureiswelcomewhenitdoes;materialthatismerelyaboutthesamesubject,howevercorrect,ispadding\.Testeachline:atthestepwhererunswerelosing,doesanagentthathasreaditdosomethingdifferent?

Theagentcannotedittheskillandcannotseethisprompt,soneverwritebodytexttellingitsreadertonameascriptoramendtheskill;ifascriptshouldrun,shipitandgivethecommandthatrunsit\.

Placethechangewheretheagentwillbestandingwhenitmatters\.If,atthatpoint,theagentcouldnothaveknownitwasabouttogowrong,noaddedsentencewillmoveit\-\-whatchangestherunissomethingthatmakesthewrongnessobservablethere:afileitrunswhoseoutputitmustreadbeforegoingon\.

Carrythedirectionoutinfull:afilenamedmeansthefileshippedandnamedatthepointtheagentrunsit;astepnamedmeansthestepplacedwheretheagentwillreachit;aremovalnamedmeansitisgone\.Beforereturning,readthefileastheagentwill:everypath,commandandnameinitmustexist;nothingyouaddedmaycontradictwhatyoukept;ifyouwouldnotfollowityourself,itisnotdone\.

ONEEDITISONEINTENT,notonefile\.Ifthechangeyoudecideonneedsmorethanoneparttoexistatall\-\-anewscriptisunreachableuntilasentencenamesit,amovedsectionisdanglinguntilitspointermoves\-\-thosepartsAREtheedit,andshippingthemtogetherisnottwoedits\.What"one"forbidsistwounrelatedintentsinthesamechild,becausethennomeasurementcansaywhichofthempaid\.

RETURNTHERESTOFTHEFILEBYTEFORBYTE\.YouareaskedforthewholeSKILL\.mdbecausethatistheonlywaytoreceiveachange,notbecausethewholefileisinscope\.Linesyouarenoteditingcomebackexactlyastheyarrived\-\-sameboldmarkers,samequotecharacters,samedashes,samewrapping\.Retypingthemcostssomethingreal:thereaderofthislineagecannotseewhichlinescarrytheintent,theduplicatecheckreadsalargechangewheretherewasasmallone,andthemeasurementthatfollowscannotattributewhatitmeasures\.

Rules:theresponsestartswiththecompletenewSKILL\.md,beginningwith\-\-\-\.frontmattername:mustbeexactly"trivy\-offline\-vulnerability\-scanning"\.ThedescriptionistheROUTINGLAYER:everyskill’sdescriptionsitsintheagent’scontextpermanently,whilebodiesareopenedonlywhenadescriptionwinstheagent’sattention\.Thislibraryholds3skills,soroutingpressureislow\-\-thereislittlesiblingoverlaptoremoveandtheagenthasfewdescriptionstochoosebetween,sothebody’sprocedureisusuallywheretheleverageis\.WhichlayerTHISeditactsonissettledbythedirection,notbythattendency\.Neverwidenthedescriptionintoaclaimthebodycannotserve\.ASCIIonly,noemoji,atmost315lines\(thecurrentversionis243\);donotdeletematerialgratuitously\-\-butmaterialthedirectionorproblemidentifiesasunusedormisleadingmaygo\.Preservewhatstillservesthetask\.

<<<TASK\-INSTRUCTIONsoftware\-dependency\-audit

\#\#thetaskinstructionforsoftware\-dependency\-audit\-\-theagentreadsthisinfullateveryrun

Theagentalreadyhaseverylineofit\.Repeatinganyofitintheskillchangesnothingtheagentdoes;contradictingit\(adifferentoutputpath,adifferentlibraryversion\)coststherun\.

Youareasoftwaresecurityengineer\.Givenadependencyfile,youneedtoperformasecurityaudittoidentifyvulnerabilitiesinthird\-partydependencies\.

Thedependencyfileisgivenin‘/root/package\-lock\.json‘\.

Youcanuseofflinetoolsordatabase\.

OnlydetectthevulnerabilitieswithseveritylevelsofHIGHandCRITICAL\.

Foreachvulnerability,collectthefollowinginformation:

\-Packagename

\-Installedversion

\-CVEID

\-Severitylevel

\-CVSSscore\(e\.g\.fromNVD,GHSA,orRedHat\)

\-Fixedversion\(ifavailable;ifnotavailable,leaveitN/A\)

\-Vulnerabilitytitle/description

\-ReferenceURL

Writetheresultsto‘/root/security\_audit\.csv‘withthefollowingcolumnsas

‘Package,Version,CVE\_ID,Severity,CVSS\_Score,Fixed\_Version,Title,Url‘

\>\>\>END\-TASK\-INSTRUCTION

\#\#Measuredroutingfact\(15rollout\(s\);openedin14of15\)

14of15previousrollout\(s\)openedtrivy\-offline\-vulnerability\-scanning/SKILL\.md,andnoneofthemopenedanyfileunderitsreferences/,scripts/orassets/\.

\#\#Triggersurface\(thedescriptiondecideswhetherthisbodyiseveropened\)

Thetask’sownvocabularyis:audit,dependencies,dependency,engineer,file,given,identify,party,perform,security,software,third,vulnerabilities\.AdescriptionthatDISCLAIMSoneofthesestopstheskillfiringonthetaskitexistsfor;excludingatasktheskilldoeshelpcostsscoredirectly,andexcludingoneitdoesnothelpgainsnothingunlessasiblingskillisbeingconfusedwithit\.

trivy\-offline\-vulnerability\-scanningsharesatmost5distinctiveterm\(s\)withanyofthe2otherskill\(s\)inthislibrary\.

\#\#Shippinganexecutablehelper

YoumayappendoneormorefilesaftertheSKILL\.md,eachintroducedbyitsownline:

===SCRIPTscripts/<filename\>===

<completefilecontent\>

Gates,eachenforcedmechanically:

\-atmost4files,allunderscripts/,40000bytestotal

\-theSKILL\.mdmustNAMEeveryfileyouship,atthepointintheprocedurewheretheagentshouldrunit;anunnamedfileisunreachableandthechildisrejected

\-everyfileisEXECUTEDinsidethistask’sowncontainerbeforethechildisaccepted:itsmodulelevelmustruntocompletionanditmustsurvivebeingrunwithnoarguments\-\-importonlywhatthatcontainerimagealreadyhas,andstopcleanlywheninputsaremissingratherthanassumingthem

\-adirectionthatnamesanexecutable,answeredwithprosealone,hasnotmadetheedit;deliverthefile

AfiletheskillALREADYshipsischangedbyre\-emittingthewholefilewithyourchangeinit;keepeverythingthatstillworks\-\-thediffagainsttheinheritedversionisrecorded\.Changeoneonlywherethedirectionsaysthatfileiswhatiswrong\.

Whenthedirectiondoesnotnameafileyoumaystillshipone,andwherethecauseisthatrunssubmitsomethingunchecked,youusuallyshould:aninstructioncanonlysaywhattoDO,whileachecktheagentrunscansaywhetheritwasDONE,andacheckisfollowedeverytimewhileasentenceisfollowedaswellasithappenstobe\.WhatafileisFOR:toderiveavaluetheagenthasnowaytolookup,ortoletitOBSERVEaresultitotherwisesubmitsunchecked\.Notfordecoration\-\-shiponethatactsonthecause,notonethatdemonstratesthecapability\.Thefileandthesentencethatmakesitreachableareoneedit\(seeONEEDITISONEINTENTabove\)\.

\#\#Currentversion

\[CURRENTSKILLDOCUMENT\]

\#\#Whattooutput,nowthatyouhavereadtrivy\-offline\-vulnerability\-scanning

\-theCOMPLETEnewSKILL\.md,startingwith\-\-\-

\-youMAYappendfilesafterit,eachintroducedbyaline===SCRIPTscripts/<filename\>===,andshouldwherethedirectionorthecauseisthatrunssubmitsomethingunchecked

\-makethisoneeditandnoother:Rewritethescaninstructionstokeepthisauditon‘/root/package\-lock\.json‘withthedefaultofflineproduction\-dependencyscopeandexplicitlyreject‘\-\-include\-dev\-deps‘\.

\-everylineyouaddhastochangewhattheagentDOESatthestepwhererunswerelosing;materialthatreadsliketherestofthisfilebutdoesnotactonthatispadding